Insight and algorithmic clustering for automated synthesis (2024)

The present application is a:

Continuation of U.S. patent application Ser. No. 15/470,298, filed Mar. 27, 2017, now U.S. Pat. No. 10,318,503, issued Jun. 11, 2019, which is a

Continuation of U.S. patent application Ser. No. 15/148,877, filed May 6, 2016, now U.S. Pat. No. 9,607,023, issued Mar. 28, 2017, which is a

Continuation of U.S. patent application Ser. No. 13/826,338, filed Mar. 14, 2013, now U.S. Pat. No. 9,336,302, issued May 10, 2016, which is a

Non-provisional and claims benefit of priority under 35 U.S.C. § 119(e) from U.S. Provisional Application No. 61/673,914, filed Jul. 20, 2012,

the entirety of which are each expressly incorporated herein by reference in their entirety.

The present invention relates to the field of domain or context-specific automated data classification or truth-seeking, and more particularly to the field of semi-supervised data clustering.

Investment Insights

Understanding the true value of assets of any kind is a critical problem in purchase decisions. Often, the available markets are inefficient at finding the true value (appropriate risk adjusted return) of the assets. As such, it is difficult for an investor or potential investor to make, monitor and communicate decisions around comparison of the value of two assets, comparing the value of two different asset types, and comparing the value of an asset to the market of all such assets.

Most types of assets have a price (as determined by what you pay for it) and a value (what it is actually worth). If you paid the right price, you would realize the expected value. Value is reflective of the price in relation to the risk-adjusted return. In poorly valued items, the price is subject to arbitrage. As the market becomes more efficient, the price and value should converge. Duration is the time horizon between when an interest in an asset is acquired as a primary and when sale, transfer, or the like realizes the complete value, e.g., in a completely measurable asset such as treasury cash, e.g., by sale, transfer, or the like. As the duration is increased, a secondary market or exchange may be available to trade the asset, and speculative investors may take part seeking to profit off changes in risk or reward over time, and often not focusing on the true value.

While there is much emphasis on data correctness, using incomplete data can also lead to faulty analysis. Overall completeness consists of the various completeness factors being monitored, as well as the timeliness of the data. Coverage refers to the percent of the total universe covered by a dataset and the expanse (or breadth) of the coverage. Proper coverage metrics must address the issue of whether a particular data set is indeed a representative sample of the space, such that analysis run on the data set gives results that are consistent with the same analysis on the full universe. Correctness, completeness and coverage are a function of the statistical confidence interval that is required by the user on the results. User may trade-off confidence to increase other parameters.

When investing in assets, each with its own risk and reward profile, it is important for an investor to understand what risks (including pricing risk) are being incurred, what return on investment is to be expected, what term/duration of investment is considered, and how a hypothetical efficient market would value the asset. In terms of market valuation (especially in an inefficient market that does not correctly price assets), it is important to compare an asset to three potential different valuation anchors, its “peers”, or assets which have similar risks vs. returns to evaluate which provides superior value (or explained and accounted for differences in risk and reward profiles), the overall market benchmark (i.e. the largest universe of peers that represent the appropriate market) and potentially a broader category of peers from other markets (or clusters). In some cases, the identification of the peer group may be difficult, and not readily amenable to accurate fully automated determination. Humans, however, may have insights that can resolve this issue.

Some users may have superior insights into understanding the price, risks and rewards, and thus a better assessment of the value or risk adjusted return than others. Investors seek benefit of those with superior insights as advisors. Those with superior insights can recognize patterns that others might not see, and classify data and draw abstractions and conclusions differently than others. However, prospectively determining which advisor(s) to rely on in terms of superior return on investment while incurring acceptable risk remains an unresolved problem.

Data Clustering

Data clustering is a process of grouping together data points having common characteristics. In automated processes, a cost function or distance function is defined, and data is classified is belonging to various clusters by making decisions about its relationship to the various defined clusters (or automatically defined clusters) in accordance with the cost function or distance function. Therefore, the clustering problem is an automated decision-making problem. The science of clustering is well established, and various different paradigms are available. After the cost or distance function is defined and formulated as clustering criteria, the clustering process becomes one of optimization according to an optimization process, which itself may be imperfect or provide different optimized results in dependence on the particular optimization employed. For large data sets, a complete evaluation of a single optimum state may be infeasible, and therefore the optimization process subject to error, bias, ambiguity, or other known artifacts.

In some cases, the distribution of data is continuous, and the cluster boundaries sensitive to subjective considerations or have particular sensitivity to the aspects and characteristics of the clustering technology employed. In contrast, in other cases, the inclusion of data within a particular cluster is relatively insensitive to the clustering methodology. Likewise, in some cases, the use of the clustering results focuses on the marginal data, that is, the quality of the clustering is a critical factor in the use of the system.

The ultimate goal of clustering is to provide users with meaningful insights from the original data, so that they can effectively solve the problems encountered. Clustering acts to effectively reduce the dimensionality of a data set by treating each cluster as a degree of freedom, with a distance from a centroid or other characteristic exemplar of the set. In a non-hybrid system, the distance is a scalar, while in systems that retain some flexibility at the cost of complexity, the distance itself may be a vector. Thus, a data set with 10,000 data points, potentially has 10,000 degrees of freedom, that is, each data point represents the centroid of its own cluster. However, if it is clustered into 100 groups of 100 data points, the degrees of freedom is reduced to 100, with the remaining differences expressed as a distance from the cluster definition. Cluster analysis groups data objects based on information in or about the data that describes the objects and their relationships. The goal is that the objects within a group be similar (or related) to one another and different from (or unrelated to) the objects in other groups. The greater the similarity (or hom*ogeneity) within a group and the greater the difference between groups, the “better” or more distinct is the clustering.

In some cases, the dimensionality may be reduced to one, in which case all of the dimensional variety of the data set is reduced to a distance according to a distance function. This distance function may be useful, since it permits dimensionless comparison of the entire data set, and allows a user to modify the distance function to meet various constraints. Likewise, in certain types of clustering, the distance functions for each cluster may be defined independently, and then applied to the entire data set. In other types of clustering, the distance function is defined for the entire data set, and is not (or cannot readily be) tweaked for each cluster. Similarly, feasible clustering algorithms for large data sets preferably do not have interactive distance functions in which the distance function itself changes depending on the data. Many clustering processes are iterative, and as such produce a putative clustering of the data, and then seek to produce a better clustering, and when a better clustering is found, making that the putative clustering. However, in complex data sets, there are relationships between data points such that a cost or penalty (or reward) is incurred if data points are clustered in a certain way. Thus, while the clustering algorithm may split data points which have an affinity (or group together data points, which have a negative affinity, the optimization becomes more difficult.

Thus, for example, a semantic database may be represented as a set of documents with words or phrases. Words may be ambiguous, such as “apple”, representing a fruit, a computer company, a record company, and a musical artist. In order to effectively use the database, the multiple meanings or contexts need to be resolved. In order to resolve the context, an automated process might be used to exploit available information for separating the meanings, i.e., clustering documents according to their context. This automated process can be difficult as the data set grows, and in some cases the available information is insufficient for accurate automated clustering. On the other hand, a human can often determine a context by making an inference, which, though subject to error or bias, may represent a most useful result regardless.

In supervised classification, the mapping from a set of input data vectors to a finite set of discrete class labels is modeled in terms of some mathematical function including a vector of adjustable parameters. The values of these adjustable parameters are determined (optimized) by an inductive learning algorithm (also termed inducer), whose aim is to minimize an empirical risk function on a finite data set of input. When the inducer reaches convergence or terminates, an induced classifier is generated. In unsupervised classification, called clustering or exploratory data analysis, no labeled data are available. The goal of clustering is to separate a finite unlabeled data set into a finite and discrete set of “natural,” hidden data structures, rather than provide an accurate characterization of unobserved samples generated from the same probability distribution. In semi-supervised classification, a portion of the data are labeled, or sparse label feedback is used during the process.

Non-predictive clustering is a subjective process in nature, seeking to ensure that the similarity between objects within a cluster is larger than the similarity between objects belonging to different clusters. Cluster analysis divides data into groups (clusters) that are meaningful, useful, or both. If meaningful groups are the goal, then the clusters should capture the “natural” structure of the data. In some cases, however, cluster analysis is only a useful starting point for other purposes, such as data summarization. However, this often begs the question, especially in marginal cases; what is the natural structure of the data, and how do we know when the clustering deviates from “truth”?

Many data analysis techniques, such as regression or principal component analysis (PCA), have a time or space complexity of O(m²) or higher (where m is the number of objects), and thus, are not practical for large data sets. However, instead of applying the algorithm to the entire data set, it can be applied to a reduced data set consisting only of cluster prototypes. Depending on the type of analysis, the number of prototypes, and the accuracy with which the prototypes represent the data, the results can be comparable to those that would have been obtained if all the data could have been used. The entire data set may then be assigned to the clusters based on a distance function.

Clustering algorithms partition data into a certain number of clusters (groups, subsets, or categories). Important considerations include feature selection or extraction (choosing distinguishing or important features, and only such features); Clustering algorithm design or selection (accuracy and precision with respect to the intended use of the classification result; feasibility and computational cost; etc.); and to the extent different from the clustering criterion, optimization algorithm design or selection.

Finding nearest neighbors can require computing the pairwise distance between all points. However, clusters and their cluster prototypes might be found more efficiently. Assuming that the clustering distance metric reasonably includes close points, and excludes far points, then the neighbor analysis may be limited to members of nearby clusters, thus reducing the complexity of the computation.

There are generally three types of clustering structures, known as partitional clustering, hierarchical clustering, and individual clusters. The most commonly discussed distinction among different types of clusterings is whether the set of clusters is nested or unnested, or in more traditional terminology, hierarchical or partitional. A partitional clustering is simply a division of the set of data objects into non-overlapping subsets (clusters) such that each data object is in exactly one subset. If the clusters have sub-clusters, then we obtain a hierarchical clustering, which is a set of nested clusters that are organized as a tree. Each node (cluster) in the tree (except for the leaf nodes) is the union of its children (sub-clusters), and the root of the tree is the cluster containing all the objects. Often, but not always, the leaves of the tree are singleton clusters of individual data objects. A hierarchical clustering can be viewed as a sequence of partitional clusterings and a partitional clustering can be obtained by taking any member of that sequence; i.e., by cutting the hierarchical tree at a particular level.

There are many situations in which a point could reasonably be placed in more than one cluster, and these situations are better addressed by non-exclusive clustering. In the most general sense, an overlapping or non-exclusive clustering is used to reflect the fact that an object can simultaneously belong to more than one group (class). A non-exclusive clustering is also often used when, for example, an object is “between” two or more clusters and could reasonably be assigned to any of these clusters. In a fuzzy clustering, every object belongs to every cluster with a membership weight. In other words, clusters are treated as fuzzy sets. Similarly, probabilistic clustering techniques compute the probability with which each point belongs to each cluster.

In many cases, a fuzzy or probabilistic clustering is converted to an exclusive clustering by assigning each object to the cluster in which its membership weight or probability is highest. Thus, the inter-cluster and intra-cluster distance function is symmetric. However, it is also possible to apply a different function to uniquely assign objects to a particular cluster.

A well-separated cluster is a set of objects in which each object is closer (or more similar) to every other object in the cluster than to any object not in the cluster. Sometimes a threshold is used to specify that all the objects in a cluster must be sufficiently close (or similar) to one another. The distance between any two points in different groups is larger than the distance between any two points within a group. Well-separated clusters do not need to be spherical, but can have any shape.

If the data is represented as a graph, where the nodes are objects and the links represent connections among objects, then a cluster can be defined as a connected component; i.e., a group of objects that are significantly connected to one another, but that have less connected to objects outside the group. This implies that each object in a contiguity-based cluster is closer to some other object in the cluster than to any point in a different cluster.

A density-based cluster is a dense region of objects that is surrounded by a region of low density. A density-based definition of a cluster is often employed when the clusters are irregular or intertwined, and when noise and outliers are present. DBSCAN is a density-based clustering algorithm that produces a partitional clustering, in which the number of clusters is automatically determined by the algorithm. Points in low-density regions are classified as noise and omitted; thus, DBSCAN does not produce a complete clustering.

A prototype-based cluster is a set of objects in which each object is closer (more similar) to the prototype that defines the cluster than to the prototype of any other cluster. For data with continuous attributes, the prototype of a cluster is often a centroid, i.e., the average (mean) of all the points in the cluster. When a centroid is not meaningful, such as when the data has categorical attributes, the prototype is often a medoid, i.e., the most representative point of a cluster. For many types of data, the prototype can be regarded as the most central point. These clusters tend to be globular. K-means is a prototype-based, partitional clustering technique that attempts to find a user-specified number of clusters (K), which are represented by their centroids. Prototype-based clustering techniques create a one-level partitioning of the data objects. There are a number of such techniques, but two of the most prominent are K-means and K-medoid. K-means defines a prototype in terms of a centroid, which is usually the mean of a group of points, and is typically applied to objects in a continuous n-dimensional space. K-medoid defines a prototype in terms of a medoid, which is the most representative point for a group of points, and can be applied to a wide range of data since it requires only a proximity measure for a pair of objects. While a centroid almost never corresponds to an actual data point, a medoid, by its definition, must be an actual data point.

In the K-means clustering technique, we first choose K initial centroids, the number of clusters desired. Each point in the data set is then assigned to the closest centroid, and each collection of points assigned to a centroid is a cluster. The centroid of each cluster is then updated based on the points assigned to the cluster. We iteratively assign points and update until convergence (no point changes clusters), or equivalently, until the centroids remain the same. For some combinations of proximity functions and types of centroids, K-means always converges to a solution; i.e., K-means reaches a state in which no points are shifting from one cluster to another, and hence, the centroids don't change. Because convergence tends to b asymptotic, the end condition may be set as a maximum change between iterations. Because of the possibility that the optimization results in a local minimum instead of a global minimum, errors may be maintained unless and until corrected. Therefore, a human assignment or reassignment of data points into classes, either as a constraint on the optimization, or as an initial condition, is possible.

To assign a point to the closest centroid, a proximity measure is required. Euclidean (L2) distance is often used for data points in Euclidean space, while cosine similarity may be more appropriate for documents. However, there may be several types of proximity measures that are appropriate for a given type of data. For example, Manhattan (L1) distance can be used for Euclidean data, while the Jaccard measure is often employed for documents. Usually, the similarity measures used for K-means are relatively simple since the algorithm repeatedly calculates the similarity of each point to each centroid, and thus complex distance functions incur computational complexity. The clustering may be computed as a statistical function, e.g., mean square error of the distance of each data point according to the distance function from the centroid. Note that the K-means may only find a local minimum, since the algorithm does not test each point for each possible centroid, and the starting presumptions may influence the outcome. The typical distance functions for documents include the Manhattan (L1) distance, Bregman divergence, Mahalanobis distance, squared Euclidean distance and cosine similarity.

An optimal clustering will be obtained as long as two initial centroids fall anywhere in a pair of clusters, since the centroids will redistribute themselves, one to each cluster. As the number of clusters increases, it is increasingly likely that at least one pair of clusters will have only one initial centroid, and because the pairs of clusters are further apart than clusters within a pair, the K-means algorithm will not redistribute the centroids between pairs of clusters, leading to a suboptimal local minimum. One effective approach is to take a sample of points and cluster them using a hierarchical clustering technique. K clusters are extracted from the hierarchical clustering, and the centroids of those clusters are used as the initial centroids. This approach often works well, but is practical only if the sample is relatively small, e.g., a few hundred to a few thousand (hierarchical clustering is expensive), and K is relatively small compared to the sample size. Other selection schemes are also available.

The space requirements for K-means are modest because only the data points and centroids are stored. Specifically, the storage required is O((m+K)ⁿ), where m is the number of points and n is the number of attributes. The time requirements for K-means are also modest-basically linear in the number of data points. In particular, the time required is O(I×K×m×n), where I is the number of iterations required for convergence. As mentioned, I is often small and can usually be safely bounded, as most changes typically occur in the first few iterations. Therefore, K-means is linear in m, the number of points, and is efficient as well as simple provided that K, the number of clusters, is significantly less than m.

Outliers can unduly influence the clusters, especially when a squared error criterion is used. However, in some clustering applications, the outliers should not be eliminated or discounted, as their appropriate inclusion may lead to important insights. In some cases, such as financial analysis, apparent outliers, e.g., unusually profitable investments, can be the most interesting points.

Hierarchical clustering techniques are a second important category of clustering methods. There are two basic approaches for generating a hierarchical clustering: Agglomerative and divisive. Agglomerative clustering merges close clusters in an initially high dimensionality space, while divisive splits large clusters. Agglomerative clustering relies upon a cluster distance, as opposed to an object distance. For example, the distance between centroids or medioids of the clusters, the closest points in two clusters, the further points in two clusters, or some average distance metric. Ward's method measures the proximity between two clusters in terms of the increase in the sum of the squares of the errors that results from merging the two clusters.

Agglomerative Hierarchical Clustering refers to clustering techniques that produce a hierarchical clustering by starting with each point as a singleton cluster and then repeatedly merging the two closest clusters until a single, all-encompassing cluster remains. Agglomerative hierarchical clustering cannot be viewed as globally optimizing an objective function. Instead, agglomerative hierarchical clustering techniques use various criteria to decide locally, at each step, which clusters should be merged (or split for divisive approaches). This approach yields clustering algorithms that avoid the difficulty of attempting to solve a hard combinatorial optimization problem. Furthermore, such approaches do not have problems with local minima or difficulties in choosing initial points. Of course, the time complexity of O(m² log m) and the space complexity of O(m²) are prohibitive in many cases. Agglomerative hierarchical clustering algorithms tend to make good local decisions about combining two clusters since they can use information about the pair-wise similarity of all points. However, once a decision is made to merge two clusters, it cannot be undone at a later time. This approach prevents a local optimization criterion from becoming a global optimization criterion.

In supervised classification, the evaluation of the resulting classification model is an integral part of the process of developing a classification model. Being able to distinguish whether there is non-random structure in the data is an important aspect of cluster validation.

Each of the following is expressly incorporated herein by reference in its entirety, for example, for its disclosure of clustering technology, applications of that technology, and implementations: 20120137182; 20120136860; 20120131701; 20120125178; 20120123279; 20120109778; 20120089341; 20120088981; 20120084251; 20120078927; 20120078906; 20120078858; 20120076372; 20120072124; 20120070452; 20120054133; 20120047098; 20120045119; 20120041955; 20120039541; 20120036096; 20120030185; 20120030165; 20120021710; 20120015841; 20120014560; 20120011135; 20120005238; 20110320396; 20110306354; 20110304619; 20110301860; 20110299765; 20110295773; 20110288890; 20110282877; 20110282828; 20110269479; 20110264432; 20110261049; 20110255748; 20110255747; 20110251081; 20110246483; 20110246409; 20110246200; 20110231414; 20110225158; 20110221767; 20110218990; 20110206246; 20110205399; 20110202540; 20110191353; 20110191283; 20110191076; 20110185234; 20110179019; 20110178965; 20110175905; 20110173189; 20110173173; 20110172501; 20110166949; 20110161205; 20110144914; 20110144480; 20110143650; 20110142318; 20110142287; 20110124525; 20110119108; 20110116690; 20110115787; 20110106801; 20110105350; 20110105340; 20110103613; 20110097001; 20110093492; 20110093482; 20110091083; 20110091074; 20110091073; 20110086349; 20110081066; 20110081056; 20110080490; 20110078144; 20110078143; 20110060717; 20110060716; 20110055192; 20110052076; 20110048731; 20110047172; 20110040192; 20110029657; 20110022599; 20110022354; 20110020779; 20110015869; 20110013840; 20110008805; 20110004578; 20110004415; 20110004115; 20110002194; 20110002028; 20100332475; 20100332474; 20100332425; 20100332242; 20100332210; 20100322525; 20100318492; 20100313157; 20100305930; 20100305868; 20100299128; 20100296748; 20100284915; 20100280987; 20100278425; 20100268512; 20100268476; 20100257092; 20100254614; 20100250527; 20100250477; 20100239147; 20100232718; 20100228731; 20100228625; 20100221722; 20100217763; 20100216660; 20100215903; 20100205213; 20100204061; 20100199186; 20100198864; 20100191722; 20100191532; 20100189333; 20100183555; 20100174985; 20100174983; 20100174982; 20100174980; 20100174979; 20100174978; 20100174977; 20100174976; 20100174732; 20100174492; 20100169025; 20100166339; 20100161590; 20100161232; 20100157340; 20100157089; 20100150453; 20100149917; 20100138894; 20100136553; 20100135597; 20100135582; 20100125594; 20100121638; 20100117978; 20100114929; 20100114928; 20100112234; 20100111396; 20100111370; 20100106713; 20100100515; 20100085358; 20100082614; 20100082367; 20100081661; 20100080439; 20100076981; 20100067745; 20100057534; 20100057399; 20100057391; 20100055678; 20100054278; 20100050260; 20100049431; 20100042563; 20100036647; 20100034422; 20100033182; 20100017487; 20100005105; 20100004923; 20100004898; 20090327185; 20090326383; 20090319526; 20090319454; 20090318815; 20090313294; 20090311786; 20090299990; 20090299822; 20090299705; 20090297048; 20090292802; 20090292695; 20090292694; 20090292482; 20090290778; 20090287689; 20090277322; 20090276705; 20090271694; 20090271424; 20090271405; 20090271404; 20090271397; 20090271363; 20090271359; 20090271246; 20090265024; 20090252046; 20090248399; 20090234876; 20090226081; 20090222430; 20090220488; 20090204609; 20090204574; 20090204333; 20090199099; 20090190798; 20090175545; 20090175544; 20090169065; 20090164192; 20090154795; 20090150340; 20090132347; 20090125916; 20090125482; 20090124512; 20090104605; 20090097728; 20090094265; 20090094233; 20090094232; 20090094231; 20090094209; 20090094208; 20090094207; 20090094021; 20090094020; 20090093717; 20090083211; 20090081645; 20090080777; 20090077093; 20090070346; 20090063537; 20090060042; 20090055257; 20090055147; 20090048841; 20090043714; 20090028441; 20090024555; 20090022472; 20090022374; 20090012766; 20090010495; 20090006378; 20080319973; 20080310005; 20080302657; 20080300875; 20080300797; 20080275671; 20080267471; 20080263088; 20080261820; 20080261516; 20080260247; 20080256093; 20080249414; 20080243839; 20080243817; 20080243816; 20080243815; 20080243638; 20080243637; 20080234977; 20080232687; 20080226151; 20080222225; 20080222075; 20080221876; 20080215510; 20080212899; 20080208855; 20080208828; 20080201397; 20080198231; 20080198160; 20080191035; 20080189306; 20080188964; 20080183546; 20080182282; 20080181479; 20080177640; 20080177538; 20080162541; 20080155335; 20080152231; 20080147655; 20080147591; 20080147441; 20080147440; 20080147438; 20080146334; 20080144943; 20080126464; 20080123940; 20080114800; 20080114756; 20080114710; 20080114564; 20080112684; 20080109288; 20080101705; 20080097820; 20080091423; 20080082426; 20080077570; 20080069437; 20080057590; 20080037872; 20080037536; 20080033658; 20080030836; 20080010605; 20080010273; 20080010272; 20080010262; 20080010045; 20080005137; 20070291958; 20070288465; 20070286489; 20070285575; 20070276723; 20070275108; 20070269804; 20070263900; 20070255707; 20070250522; 20070244768; 20070239982; 20070239741; 20070239694; 20070233711; 20070231921; 20070217676; 20070198553; 20070192063; 20070192034; 20070185946; 20070180980; 20070179784; 20070174335; 20070172803; 20070156516; 20070154931; 20070154066; 20070150443; 20070141527; 20070129991; 20070129011; 20070128573; 20070111316; 20070106405; 20070093966; 20070092905; 20070092888; 20070078846; 20070067212; 20070064627; 20070054266; 20070050708; 20070044010; 20070038612; 20070033533; 20070033521; 20070033515; 20070033292; 20070033221; 20070033214; 20070033170; 20070025637; 20070022279; 20070008905; 20070006177; 20070005556; 20070003138; 20060282425; 20060282298; 20060281473; 20060253258; 20060248141; 20060246495; 20060239338; 20060224356; 20060212337; 20060208185; 20060195415; 20060195269; 20060195204; 20060190465; 20060190191; 20060177837; 20060136589; 20060112146; 20060106816; 20060101377; 20060101060; 20060095521; 20060093208; 20060093188; 20060074924; 20060074771; 20060074621; 20060064177; 20060058592; 20060053142; 20060053129; 20060052943; 20060041414; 20060034545; 20060031219; 20060020662; 20060015630; 20060015341; 20060013482; 20050286774; 20050285937; 20050283328; 20050281291; 20050278324; 20050273319; 20050267993; 20050267992; 20050267991; 20050265331; 20050262044; 20050256413; 20050255458; 20050251882; 20050225678; 20050198575; 20050193216; 20050192768; 20050185848; 20050182570; 20050180638; 20050176057; 20050175244; 20050164273; 20050163384; 20050163373; 20050149269; 20050147303; 20050138056; 20050137806; 20050132069; 20050130230; 20050130215; 20050120105; 20050114331; 20050102305; 20050102272; 20050085436; 20050075995; 20050058336; 20050027829; 20050015376; 20050010571; 20040267774; 20040260694; 20040254901; 20040249939; 20040249789; 20040243362; 20040233987; 20040230586; 20040213461; 20040181527; 20040177069; 20040175700; 20040172225; 20040171063; 20040170318; 20040162834; 20040162647; 20040139067; 20040130546; 20040129199; 20040127777; 20040122797; 20040107205; 20040103377; 20040101198; 20040091933; 20040075656; 20040071368; 20040068332; 20040056778; 20040049517; 20040048264; 20040036716; 20040024773; 20040024758; 20040024739; 20040019574; 20040013292; 20040003005; 20040002973; 20040002954; 20030229635; 20030208488; 20030205124; 20030175720; 20030174179; 20030161500; 20030161396; 20030158842; 20030145014; 20030139851; 20030138978; 20030129660; 20030120630; 20030107768; 20030101003; 20030100996; 20030097357; 20030097356; 20030093227; 20030088563; 20030078509; 20030078494; 20030074251; 20030065661; 20030065635; 20030061249; 20030059081; 20030058339; 20030054573; 20030050923; 20030050908; 20030046253; 20030046018; 20030044062; 20030044053; 20030036093; 20030033138; 20030028564; 20030016250; 20030014191; 20030009469; 20030009333; 20020191034; 20020190198; 20020184080; 20020183966; 20020181786; 20020181711; 20020147703; 20020146175; 20020143989; 20020132479; 20020131641; 20020129038; 20020128781; 20020122587; 20020115070; 20020111966; 20020099721; 20020099675; 20020091655; 20020069218; 20020050990; 20020049740; 20020033835; 20020023061; 20020002555; 20020002550; 20020000986; 20010055019; 20010048753; 20010014868; 20010000356; U.S. Pat. Nos. 8,200,648; 8,200,506; 8,195,734; 8,195,670; 8,195,345; 8,191,783; 8,190,663; 8,190,082; 8,184,913; 8,183,050; 8,180,766; 8,180,627; 8,180,147; 8,175,896; 8,175,730; 8,175,412; 8,170,961; 8,170,306; 8,169,681; 8,169,481; 8,165,407; 8,165,406; 8,164,507; 8,150,169; 8,145,669; 8,139,838; 8,135,719; 8,135,681; 8,135,680; 8,135,679; 8,122,502; 8,122,045; 8,117,213; 8,117,204; 8,117,203; 8,117,139; 8,116,566; 8,108,931; 8,108,405; 8,108,392; 8,099,381; 8,097,469; 8,095,830; 8,095,521; 8,095,389; 8,090,729; 8,082,246; 8,077,984; 8,073,652; 8,065,316; 8,065,248; 8,055,677; 8,051,139; 8,051,082; 8,046,362; 8,041,715; 8,032,476; 8,027,977; 8,019,766; 8,015,183; 8,015,125; 8,015,124; 8,014,957; 8,014,591; 8,010,589; 8,010,466; 8,005,294; 8,000,533; 8,000,527; 7,996,369; 7,991,557; 7,979,435; 7,979,362; 7,975,039; 7,975,035; 7,970,627; 7,966,327; 7,966,225; 7,966,130; 7,962,651; 7,958,096; 7,954,090; 7,953,705; 7,953,679; 7,949,186; 7,937,349; 7,937,234; 7,933,915; 7,933,740; 7,930,189; 7,926,026; 7,917,517; 7,917,306; 7,912,734; 7,912,726; 7,912,290; 7,912,284; 7,904,303; 7,899,564; 7,894,995; 7,894,669; 7,890,512; 7,890,510; 7,890,294; 7,889,914; 7,889,679; 7,885,966; 7,882,126; 7,882,119; 7,879,620; 7,876,947; 7,873,616; 7,868,786; 7,865,456; 7,856,434; 7,849,027; 7,848,567; 7,842,874; 7,835,542; 7,831,549; 7,831,531; 7,831,325; 7,827,183; 7,827,181; 7,826,635; 7,823,055; 7,822,426; 7,813,580; 7,805,496; 7,805,443; 7,805,266; 7,801,893; 7,801,685; 7,783,249; 7,773,784; 7,767,395; 7,761,448; 7,752,208; 7,747,547; 7,747,390; 7,747,054; 7,746,534; 7,743,059; 7,739,284; 7,736,905; 7,716,148; 7,711,846; 7,707,210; 7,702,155; 7,697,785; 7,693,683; 7,689,457; 7,688,495; 7,685,090; 7,684,963; 7,679,617; 7,660,468; 7,657,379; 7,657,126; 7,657,100; 7,650,320; 7,644,090; 7,643,597; 7,639,868; 7,639,714; 7,624,337; 7,613,572; 7,610,306; 7,603,326; 7,599,917; 7,599,799; 7,590,264; 7,584,168; 7,580,682; 7,580,556; 7,574,409; 7,574,069; 7,570,213; 7,567,961; 7,565,432; 7,565,346; 7,565,251; 7,565,213; 7,562,325; 7,562,015; 7,558,425; 7,555,441; 7,552,474; 7,552,131; 7,545,978; 7,539,656; 7,529,732; 7,526,101; 7,519,227; 7,519,209; 7,516,149; 7,512,524; 7,499,916; 7,492,943; 7,487,056; 7,475,085; 7,468,730; 7,464,074; 7,458,050; 7,450,746; 7,450,122; 7,437,308; 7,428,541; 7,428,528; 7,426,301; 7,424,462; 7,418,136; 7,406,200; 7,401,087; 7,397,946; 7,395,250; 7,389,281; 7,386,426; 7,376,752; 7,369,961; 7,369,889; 7,369,680; 7,346,601; 7,337,158; 7,328,363; 7,325,201; 7,296,088; 7,296,011; 7,293,036; 7,287,019; 7,275,018; 7,272,262; 7,263,220; 7,251,648; 7,246,128; 7,246,012; 7,231,074; 7,225,397; 7,222,126; 7,221,794; 7,216,129; 7,215,786; 7,206,778; 7,202,791; 7,196,705; 7,188,055; 7,177,470; 7,174,048; 7,167,578; 7,158,970; 7,142,602; 7,139,739; 7,111,188; 7,068,723; 7,065,587; 7,065,521; 7,062,083; 7,058,650; 7,058,638; 7,054,724; 7,047,252; 7,043,463; 7,039,621; 7,039,446; 7,035,823; 7,035,431; 7,031,980; 7,031,844; 7,016,531; 7,010,520; 6,999,886; 6,993,186; 6,980,984; 6,976,016; 6,970,796; 6,968,342; 6,961,721; 6,954,756; 6,950,752; 6,915,241; 6,912,547; 6,907,380; 6,906,719; 6,904,420; 6,895,267; 6,854,096; 6,845,377; 6,841,403; 6,834,278; 6,834,266; 6,832,162; 6,826,316; 6,819,793; 6,816,848; 6,807,306; 6,804,670; 6,801,859; 6,801,645; 6,799,175; 6,797,526; 6,785,419; 6,785,409; 6,778,981; 6,778,699; 6,763,128; 6,760,701; 6,757,415; 6,751,614; 6,751,363; 6,750,859; 6,735,465; 6,735,336; 6,732,119; 6,711,585; 6,701,026; 6,700,115; 6,684,177; 6,674,905; 6,643,629; 6,636,849; 6,627,464; 6,615,205; 6,594,658; 6,592,627; 6,584,433; 6,564,197; 6,556,983; 6,539,352; 6,535,881; 6,526,389; 6,519,591; 6,505,191; 6,496,834; 6,487,554; 6,473,522; 6,470,094; 6,468,476; 6,466,695; 6,463,433; 6,453,246; 6,445,391; 6,437,796; 6,424,973; 6,424,971; 6,421,612; 6,415,046; 6,411,953; 6,400,831; 6,389,169; 6,373,485; 6,351,712; 6,331,859; 6,300,965; 6,295,514; 6,295,504; 6,295,367; 6,282,538; 6,263,334; 6,263,088; 6,249,241; 6,203,987; 6,192,364; 6,185,314; 6,140,643; 6,122,628; 6,121,969; 6,112,186; 6,100,825; 6,092,049; 6,085,151; 6,049,777; 6,041,311; 5,949,367; 5,940,833; 5,940,529; 5,926,820; 5,920,852; 5,889,523; 5,872,850; 5,813,002; 5,809,490; 5,795,727; 5,764,283; 5,748,780; 5,731,989; 5,724,571; 5,717,915; 5,710,916; 5,699,507; 5,668,897; 5,627,040; 5,625,704; 5,574,837; 5,566,078; 5,506,801; 5,497,486; 5,463,702; 5,448,684; 5,442,792; 5,327,521; 5,285,291; 5,253,307; 5,020,411; 4,965,580; 4,855,923; 4,773,093; 4,257,703; and 4,081,607.

The Reference-User

The present technology provides a system and method which exploits human interactions with an automated database system to derive insights about the data structures that are difficult, infeasible, or impossible to extract in a fully automated fashion, and to use these insights to accurately assess a risk adjusted value or cluster boundaries.

According to an aspect of the technology, the system monitors or polls a set of users, actively using the system or interacting with the outputs and providing inputs. The inputs may be normal usage, i.e., the user is acting in a goal directed manner, and providing inputs expressly related to the important issues, or explicit feedback, in which the user acts to correct or punish mistakes made by the automated system, and/or reward or reinforce appropriate actions.

Through automated historical and action-outcome analysis, a subset of users, called “reference-users” are identified who demonstrate superior insight into the issue or sub-issue important to the system or its users. After the reference-users are identified, their actions or inputs are then used to modify or influence the data processing, such as to provide values or cluster the data. The adaptive algorithm is also able to demote reference-users to regular users. Additionally, because reference-user status may give rise to an ability to influence markets, some degree of random promotion and demotion is employed, to lessen the incentive to exploit an actual or presumed reference-user status. Indeed, the system may employ a genetic algorithm to continuously select appropriate reference-users, possibly through injection of “spikes” or spurious information, seeking to identify users that are able to identify the spurious data, as an indication of users who intuitively understand the data model and its normal and expected range. Thus, the system is continuously or sporadically doing three things—learning from reference-users and learning who is a reference-user, requesting more granulation/tagging and using that learning to cluster/partition the dataset for the ordinary users for the most contextually relevant insight.

Often, the reference-user's insights will be used to prospectively update the analytics, such as the distance function, clustering initial conditions or constraints, or optimization. However, in some cases, the adaptivity to the reference-user will only occur after verification. That is, a reference-user will provide an input which cannot contemporaneously be verified by the automated system. That input is stored, and the correspondence of the reference-user's insight to later reality then permits a model to be derived from that reference-user which is then used prospectively. This imposes a delay in the updating of the system, but also does not reveal the reference-user's decisions immediately for use by others. Thus, in a financial system, a reference-user might wish to withhold his insights from competitors while they are competitively valuable. However, after the immediate value has passed, the algorithm can be updated to benefit all. In an investment system, often a reference-user with superior insight would prefer that others follow, since this increases liquidity in the market, giving greater freedom to the reference-user.

A key issue is that a fully automated database analysis may be defined as an NP problem and in a massive database, the problem becomes essentially infeasible. However, humans tend to be effective pattern recognition engines, and reference-users may be selected that are better than average, and capable of estimating an optimal solution to a complex problem “intuitively”, that is, without a formal and exact computation, even if computationally infeasible. As stated above, some humans are better than others at certain problems, and once these better ones are identified, their insights may be exploited to advantage.

In clustering the database, a number of options are available to define the different groups of data. One option is to define persons who have a relationship to the data. That is, instead of seeking to define the context as an objective difference between data, the subjective relationships of users to data may define the clusters. This scenario redefines the problem from determining a cluster definition as a “topic” to determining a cluster definition as an affinity to a person. Note that these clusters will be quite different in their content and relationships, and thus have different application.

Optimal clustering is only one aspect of the use of a reference-user. More generally, the reference-user is a user that demonstrates uncommon insight with respect to an issue. For example, insight may help find clusters of data that tend to gravitate toward or away from each other and form clusters of similarity or boundaries. Clustering is at the heart of human pattern recognition, and involves information abstraction, classification and discrimination.

Thus, according to the present technology, we consider a system having a network of “users”, which may be ordinary human users that simply require the computer to synthesize some insight from a large dataset, as well as “reference-users” who help the computer refine and set context in the dataset. While the designation of user and reference-user partitions the space of users. The process of selecting who is a user and who is a reference-user is automated and the designations may not be persistent, i.e., the computer is continually re-evaluating who is a user and who is a reference-user based on how they interact with the system.

From a database user's perspective, a query should be simple, e.g., “natural language”, and not require a specific knowledge of the data structures within the database or a complete knowledge of the data structures being searched. In other words, the user should not have to know the structure of database before the query result is obtained. The result should preferably include relevant responses sorted or organized according to relationship with the query. In other cases, the sorting or ranking may be according to different criteria. Much as the clustering problem employs a distance function, the reporting problem also employs a ranking or information presentation prioritization function. Indeed, the outputs may be clustered either according to the clustering of the source database, or the responses to a query may be clustered upon demand.

In some cases, a user wishes only results with high relevance, while in other cases, a user may wish to see a ranked list which extends to low relevance/low yield results. A list, however, is not the only way to organize results, and, in terms of visual outputs, these may be provided as maps (see U.S. Pat. No. 7,113,958 (Three-dimensional display of document set); U.S. Pat. No. 6,584,220 (Three-dimensional display of document set); U.S. Pat. No. 6,484,168 (System for information discovery); U.S. Pat. No. 6,772,170 (System and method for interpreting document contents), each of which is expressly incorporated herein by reference), three or higher dimensional representations, or other organizations and presentations of the data. Thus, the distinction between the query or input processing, to access selected information from a database, and the presentation or output processing, to present the data to a user, is important. In some cases, these two functions are interactive, and for example, a context may be used preferentially during presentation rather than selection.

According to one embodiment of the system and method according to the present technology, a reference-user is employed periodically to normalize a data distribution, based on the reference-user's insights. This normalization acts as a correction to an automated algorithm, and the normalization information received from the reference-user tunes the algorithm, which, for example, represents distance function or partition (clustering) function. In effect the reference-users train the system when they unconsciously partition elements from the cluster based on their instincts.

The system does not have to be continuously trained by the reference-user or act to continuously reselect reference-users. The training is important only when the divergence between what the system reports as insight on a self-similar cluster and what the dominant set of users consider to be an insight, becomes unacceptably large. When this divergence becomes unacceptably large for the remaining users in the network, then the reference-user training is invoked and the system learns from the reference-user. If the divergence corrects, the systems stops retraining and continues as before. However, if the divergence does not, then the system reselects the reference-user and then retrains. Once again if the divergence corrects, the system continues as before. However, if it does not, the system then flags the needs for more data by requesting additional meta-tagging of the content.

Thus, the system is continuously doing 3 things (a) learning from reference-users; (b) learning who is a reference-user; and (c) requesting more granulation and using that learning to cluster/partition the dataset for the ordinary users for the most contextually relevant insight.

Context-Based Reference-Users

Clustering of massive databases poses a number of problems. For example, the computational complexity of some algorithms is sufficiently high that clustering cannot be updated in real time. Further, an inherent challenge in automated clustering comes from realizing that a machine may have no context, or that solution of the clustering problem could be significantly facilitated by determination and exploitation of context information. Thus, superior clustering in various cases requires the establishment of context by some means to facilitate filtering of the clusters by the computer algorithm. Another aspect of this problem is that the bases for the clustering may be determined ad hoc, or the relevant distinctions available with information provided at the time of consideration.

Context can be assumed if the insight required, and dataset to be algorithmically operated on, is small and specialized enough. Unfortunately, in very high dimensionality databases, such as Google's semantic database of the web and related data feeds, the resulting number of clusters may be extraordinarily high, and as a result the clustered database may be of little use without further distinctions being drawn. For example, the Google search engine requires a query, and even then returns result based on multiple undistinguished contexts, leading to a potentially large proportion of irrelevant responses. Likewise, simplifying presumptions made to reduce complexity may eliminate the very distinctions that are required for a particular circ*mstance.

While computers have computational power for performing standard algorithmic calculations, humans have the ability to immediately judge context. Humans do this contextual mapping by looking for similarity in networks, similarity in knowledge sets and similarity in skills. Thus, an automated way of identifying how to elicit that human “secret sauce” around context, will significantly speed up the computer's ability to partition the space into proper contextually relevant clusters.

Implicit in natural language searching and “relevance” to a query is the concept of “context”. A Boolean text search does not benefit from knowledge of language structures and latent ambiguities, and thus will typically deliver results that are contextually irrelevant but meet the Boolean query criteria. On the other hand, natural language search technologies and other unstructured search systems can benefit from context, though often determining this context requires an inference. Alternately, a user can define a context, for example by limiting himself or herself to a special purpose database or other limitation. A user can also seek to explicitly indicate the context, assuming that the user is aware of the different contexts to be distinguished. However, it is often necessary to query the database before determining the clustering of responsive records, and then obtaining feedback from a user to define the context and therefore focus on respective clusters. However, in some cases the “context” that might be derived from an automated clustering of records defies semantic description, thus requiring a “clustering by example” feedback/training of the system, or other type of non-semantic guidance of the system, and which might incur a much larger effort than most users would voluntarily endure, and perhaps incur more effort and/or higher costs than simply accepting both relevant and irrelevant information in response to the query and distinguishing these manually.

The present technology proposes a solution to this problem by designating “reference-users”, that is, either the reference-user has previously indicated or proven a superior ability to operate in a certain context, or otherwise represent the context by consistency and reliability. The user context may be determined in various ways, but in the case of persistent contexts, a user profile may be developed, and a reference-user selected with whom the user has some affinity, i.e., overlapping or correlated characteristics. There are multiple ways to designate the reference-user—the system designates the reference-user based on filtering a set of users to which reference-user best represents the responses and preferences of the set. This designation of reference-user comes from affinity, which could be network-affinity (users that are closely connected in the network in that context), knowledge-affinity (users that have superior expertise in that context), or skill-affinity (users possessing specialized skills in that context).

It is noted that the reference-user is discussed as an actual single human user, but may be a hybrid of multiple users, machine assisted humans, or even paired guides.

The problem of defining the context of a user is then translated to the problem of finding a suitable reference-user or set of reference-users. In fact, the set of reference-users for a given user may have a high consistency, and as known in the field of social networking. That is, assuming that the word “friend” is properly defined, the universe of contexts for a user may be initially estimated by the contexts of interest to his or her identified friends. Such an estimation technology is best exploited in situations where error is tolerable, and where leakage of user-specific contexts is acceptable.

In some cases, the reference-user sought is one with superior insights (rather than exemplary insights), that is, the reference-user is “better” than a normal user, and as such leads the other users. This is appropriate where an economic quality function is available, and the maximization of that function does not require significant compromise. This type of system has a well-defined and generally invariant definition of “best”, especially when an economic cost-benefit function can be defined and readily adopted.

In other cases, the reference-user should be the epitome of the class, and thus not quantitatively deviant from the group as a whole. In such a case, the user with the “best” insight might be considered statistically deviant from the mean, and therefore not a good choice for designation as reference-user.

For example, in a scientific literature database, an “expert” in a field may be designated as a reference-user, and the context associated with that expert representing the field of expertise. A database so organized would cluster the documents in the database around different spheres of expertise; the narrower the expertise of the designated expert reference-user, the higher the quality of distinctions which may be drawn from other knowledge domains.

In contrast, a general purpose database such as Google may be used by fifth graders. The clustering of information based on expertise may lead to low ranking of documents appropriate for that type of user, and high ranking of documents which are either incomprehensible to the user, or lacking relevance. Thus, the goal in a general purpose system is to identify reference-users who are similarly situated to the actual user, and therefore whose usage correlates with the intended or supposed use by the user.

Indeed, these two different types of reference-users may both be used in parallel, though because they are not self-consistent as each represent “context”, these should be treated as independent or semi-independent factors.

The “expert” reference-user, for example, may be of use to the fifth grader; the reference-user profile can be used to distinguish various contexts at high levels, which can then be used to rank documents at the appropriate level for the user. The epitome reference-user may be useful to the technical user; some relevant documents may be outside the experience or sphere of the expert reference-user, and a more common reference-user may provide useful insights for ranking or segregating those documents. By pairing the expert and the epitome, a comparison of the results may also be of use, especially in ranking the results in terms of degree of specialization.

It may be useful to explicitly receive user profile information, or inferentially derive such information, in order to identify context. In addition to analyzing content associated with user actions, the speed, duration, and latency of user actions may be analyzed, as well as the range of contexts, and usage of content.

As a final note on the form of interaction of the reference user with the data, in the typical case, we assume that the reference user can choose how they filter, cluster and view the data set. Thus, in their “view”, a reference user may choose to subtract points they wish to view, or add points they wish to “view”. This process does not change the dataset itself, but merely changes the way the reference user chooses to view the dataset. It changes the filter and is merely reflective of their context.

Objects

It is therefore an object to provide a decision support system, comprising a user input port configured to receive user inputs comprising at least one user criterion and at least one user input tuning parameter representing user tradeoff preferences for producing an output from a system which selectively produces an output of tagged data in dependence on the at least one user criterion, the at least one user input tuning parameter, and a distance function; a reference-user input configured to receive at least one reference-user input parameter representing the at least one reference-user's analysis of the tagged data and the corresponding user inputs, to adapt the distance function in accordance with the reference-user inputs as a feedback signal, wherein the reference-user acts to optimize the distance function based on the user inputs and the output, and on at least one reference-user inference; and an information repository configured to store the tagged data.

It is a further object to provide a decision support system, comprising a user input port configured to receive user inputs comprising at least one user criterion and at least one user input tuning parameter representing user tradeoff preferences for producing an output from a system which selectively produces an output of tagged data in dependence on the at least one user criterion, the at least one user input tuning parameter, and a distance function; a reference-user agent configured to receive at least one reference-user input parameter representing the at least one reference-user's analysis of the tagged data and the corresponding user inputs, to adapt the distance function in accordance with the user inputs as a feedback signal, wherein the reference-user agent acts to optimize the distance function based on the user inputs and the output, and on at least one reference-user inference derived from at least one human user selected from a plurality of possible users based on an accuracy of selection according to an objective criterion; and an information repository configured to store the tagged data.

It is a still further object to provide a decision support method, comprising receiving user inputs comprising at least one user criterion, and at least one user input tuning parameter representing user tradeoff preferences for producing an output; selectively producing an output of tagged data from a clustered database in dependence on the at least one user criterion, the at least one user input tuning parameter, and a distance function; receiving at least one reference-user input parameter representing the at least one reference-user's analysis of the tagged data and the corresponding user inputs, to adapt the distance function in accordance with the reference-user inputs as a feedback signal; and clustering the database in dependence on at least the distance function, wherein the reference-user acts to optimize the distance function based on the user inputs and the output, and on at least one reference-user inference.

The clustering may be automatically performed by a processor. The database may receive new data. The distance function may be applied to cluster the database including the new data before the at least one reference-user input parameter is received. The tagged data may comprise a valuation or rating. The distance function may be adaptive to new data. The reference-user inference may represent at least one of a valuation and a validation. The user input tuning parameter may comprise a dimensionless quantitative variable that impacts a plurality of hidden dimensions. The hidden dimensions may comprise at least one of completeness, timeliness, correctness, coverage, and confidence. The user input tuning parameter may balance completeness and correctness of the tagged data in the output.

Another object provides an information access method, comprising receiving a semantic user input comprising an indication of interest in information; determining a context of the user distinctly from the semantic user input comprising an indication of interest in information; producing an output of at least tagged data from a clustered database in dependence on at least the user input, the determined context, and a distance function; monitoring a user interaction with the output; and modifying the distance function in dependence on at least the monitored user interaction.

The method may further comprise selecting at least one commercial advertisem*nt extrinsic to the tagged data from the clustered database for presentation to the user, in dependence on at least: at least one of the semantic user input and the output of tagged data, and the determined context. The selecting may be further dependent on the distance function. The monitoring may comprises monitoring a user interaction with the at least one commercial advertisem*nt, wherein the commercial advertisem*nt is selected in dependence on the distance function, and the distance function is modified based on the user interaction with a selected advertisem*nt. The method may further comprise reclustering the database in dependence on the modified distance function. The method may further comprise classifying a plurality of users, and distinguishing between difference classes of users with respect to the selection and modifying of respective ones of a plurality of distance functions. The method may further comprise determining at least one reference-user from a set of users, based on at least one fitness criterion, and selectively modifying the distance function dependent on a reference-user input in preference to a non-reference-user input. A user input is associated with a respective reference-user in dependence on the context.

Another object provides an information processing method, comprising: clustering a database comprising a plurality of information records according to semantic information contained therein, wherein information may be classified in a plurality of different clusters in dependence on a context, such that a common semantic query to the database yields different outputs over a range of contexts; producing an output identifying information records from the database in dependence on at least a user semantic input, and a distance function; receiving user feedback; and modifying at least one distance function in dependence on the user feedback.

The method may further comprise determining a contextual ambiguity from the user semantic input, soliciting contextual ambiguity resolution information from the user, and producing a follow-up output identifying information records from the database in dependence on at least a user semantic input, the contextual ambiguity resolution information, and at least one distance function selected from a plurality of available distance functions in dependence on the contextual ambiguity resolution information. The method may further comprise selecting at least one commercial advertisem*nt extrinsic to the information records in the database for presentation to the user, in dependence on at least: the user semantic input, and the contextual ambiguity resolution information. The selecting may be further dependent on at least one distance function. The method may further comprise selecting at least one commercial advertisem*nt extrinsic to the information records in the database for presentation to the user, in dependence on at least: the user semantic input, and the distance function. The monitoring may comprise monitoring a user interaction with at least one commercial advertisem*nt presented to the user as part of the output, wherein the commercial advertisem*nt is selected in dependence on at least one distance function, and the at least one distance function is modified based on the user interaction with at least one selected advertisem*nt. The method may further comprise reclustering the database in dependence on the at least one modified distance function. The method may further comprise classifying a plurality of users, and distinguishing between difference classes of users with respect to the selection and modifying of respective ones of a plurality of distance functions. The method may further comprise assigning a reference-user status to at least one user within a set of users, based on at least one fitness criterion, and selectively weighting a user contribution to a modification of a respective distance function dependent on the reference-user status of the respective user. The reference-user status may be assigned with respect to a context, and a user input is associated with a respective distance function in dependence on the context.

FIG. 1 is a flowchart according to a first embodiment of the technology;

FIG. 2 is a flowchart according to a second embodiment of the technology;

FIG. 3 is a flowchart according to a third embodiment of the technology;

FIG. 4 is a block diagram of a traditional computing system architecture; and

FIG. 5 is a flowchart according to a fourth embodiment of the technology.

Search Engine

The reference-user is exploited in various ways. In a very uncommon scenario, the reference-user directly and in real time guides a search result. That is, the reference-user acts as expert assistance for the user, a sort of reference librarian. The abilities of the reference-user are directly exploited, but this is expensive, non-scalable, and may have difficulty addressing contexts that transcend a single reference-user, such as a hybrid context.

Another way to exploit a reference-user is to obtain a rich profile of the reference-user based on close monitoring of the reference-user or explicit training by the reference-user of an automated system. In essence, the reference-user's essence is transferred to an artificial agent, which then emulates the reference-user. This technique is both scalable and relatively low cost, but may fail in new circ*mstances. That is, such systems may do a fair job at interpolation, but may have great difficulty extrapolating. Likewise, generalizations from the reference-user profile may be unhelpful, especially if the generalization transcends the reference-user scope.

A further way to exploit the reference-user is to proactively procure a decision from the reference-user, based on his or her inferences. These decisions may be focused on defining cluster boundaries, which may be tuned, for example, at the distance function, clustering criteria, or optimization level, or imposed as an external constraint by direct classification (or reclassification) of one or more objects. Thus, as part of a voluntary process, reference-users or potential reference-users may be requested to provide feedback, or usage monitored for inferential feedback, to determine classifications (clustering) and a distance function representing quantitatively how much a given item corresponds to the putative classification. The clustering properties thus derived need not be used in a “social” context; that is, a user may be able to identify the context as a cluster, without direct or indirect reference to the reference-user responsible or partially responsible for the clustering. Therefore, a kind of collaborative filter may be implemented, to identify the appropriate reference-user or affinity group, and thereafter exploit the identified reference-user or affinity group in various ways, as may be known, or described herein. In some cases, the distance function may have separate components for a value of proper classification and a cost of misclassification. For example, in a commercial database, the retrieval cost is expensive, so there may be a bias against inclusion of objects where the relevance is doubtful, as compared to objects with equal distance, but more assured relevance.

This may be especially important where the reference-users have multiple roles, or where all roles are not relevant to a context. Thus, in a social relationship system, the reference-user as a whole defines the context (which may significantly overlap for different reference-users), while in a clustered system, all that matters is the definition of cluster boundaries and/or distance function, and the problem of selecting a cluster is different than selecting a high affinity reference-user. However, not all context determination problems are difficult, and therefore other statistical or artificial intelligence technologies may be employed.

In some cases, a pre-result may be provided to a user, which requests user input to select a particular context or contexts. This technique is disfavored in a Google-type search engine, since typically, the user seeks high quality results in the first response, and consistency of search results upon repetition. On the other hand, if a user submits a query, the initial response may be context-free (or multi-context). However, the user may then be given various options for explicit or implicit feedback, such that the ranking of results changes with each new page presented by the search engine. This feedback is a natural way to receive input for defining reference-users and for obtaining inferences from reference-users or potential reference-users. In addition, there is typically insufficient time between submission of an initial search and when the initial response is expected, in order to perform a computationally complex ad hoc clustering or other document analysis. However, the delay between the initial response (first download) from a query and subsequent responses (downloads/page refreshes) may be sufficient to perform complex analytics on the subset of responsive documents. Thus, according to one aspect, a database interface is provided that implements an adaptive user interface based on feedback received. In some cases, the feedback and context definition may be persistent, but in others, the context will be used only for the immediately subsequent interactions.

It is noted that feedback from a reference-user in a Google type search engine may be derived by monitoring click-throughs from the search results. A reference-user would presumably be able to filter useful results from those of limited value. The subset of results selected by the reference-user represents a cluster, which can then be used as an exemplar for updating the clustering algorithm for future searches within the cluster domain for which the reference-user is associated.

Thus, the first response from a database may be without defined context, or even specifically designed to elicit a definition of the context from the user. The second response may benefit from analytics as well as explicit or implicit feedback from the user to define the context and/or cluster identification. In a typical massive database, results and partial results are cached, and analytics may be performed on these cached results to determine clusters of information according to various criteria. Given a user input seeking a database response, the database may initially reveal results representing different clusters that correspond to the query. The user may than select one cluster which includes responses relevant to the context. The cluster selection is then returned to the database system, which can then deliver results appropriate for that context. Note that the clusters initially presented need not directly correspond to the identified context. For example, in a complex semantic query, the cached clusters may represent distinctions made on a subset of the query, or using a fuzzy search algorithm. Once the actual cluster including relevant responses is identified, the query may be re-executed using the formal user request, and the selected relevant responses. Typically, a massive database which provides real time responses does not have sufficient time to perform iterative database processes, while the “conversational” system can exploit user latency to perform complex analytics and recursive processes.

Interactions of Reference-Users

In designating reference-users, it is sometimes useful to also designate anti-reference-users; that is, representatives of a class or context that is undesired, or those who demonstrate poor insights. Taking Google again as an example, much of the Internet includes sex and/or adult themes, frivolous or trivial sites, and the like. However, these various elements are not universally ignored, and therefore in the same way that experts on arcane academic topics can be identified, so can “experts” on Internet spam. By identifying these “experts”, a negative affinity may be defined to help avoid undesired clusters or classes of information for a user. Thus, the reference-user does not necessarily trivialize the problem to a single cluster with a universal distance function from a common centroid for all objects. Rather, by providing multiple reference-users, the user's context can be matched with the best reference-user (or hybrid/synthetic reference-user) which results in an optimum of comprehensiveness and relevance for the given context. More generally, the user need not select a single cluster/classification/reference-user as the context, but rather a combination (e.g., linear combination) of various clusters/classifications/reference-users may be employed. As appropriate, negative weights, and required logical combinations (and, or, not, etc.) may be applied. In this way, the reference-user is not necessarily an exclusive group with extraordinary capabilities, though in many cases, those types of reference-users are particularly valued.

This technology therefore has application as a filter against inappropriate content, which may be able to implement fine distinctions between acceptable and unacceptable content. In particular, an automated filter which is not guided by human decisions may have difficulty distinguishing “p*rnography” from “erotica”, while (according to Justice Potter Stewart), a reasonable human can make this distinction intuitively. Thus, at risk of requiring the reference-users to actually behold the p*rnography in order to classify it, the distinctions may be finely drawn based on the human inference; note that the reference-user is typically monitored during normal activities, and not required to perform any particular activity. This example also raises the social network issue; since p*rnography is subject to community standards, the reference-user selected for this clustering/classification must be representative of the same community as is relevant to the user, and therefore the same data may be subject to a plurality of clusterings and distance functions. Similar distinctions may be drawn in various contexts—Darwinian evolutionists vs. creationists; conservatives vs. liberals; etc. The context may thus be independent of the database, and for example relevant to an ideology of the user.

Assessments of Users

The present technology also provides education and assessments. That is, a normal user may be educated according to the insights of a reference-user, and the ability of a user to classify similarly to an exemplary reference-user may be assessed. These technologies may of course be integrated with other educational and assessment technologies.

Reference—Users in Asset Analysis

In the system and method according to the present technology, as applied to investment data, a reference-user architecture is useful for determining peer groups among different funds, managers, segments. In this case, the goal is to select a reference-user who has demonstrated exemplary past performance at the task, and thus who likely has better “insight” into the investment quality. The reference-user(s) in this case are selected adaptively based on performance, and thus if a prior reference-user loses reliability, he is demoted. In general, the reference-user is not publicly designated, and has no knowledge that he or she serves as a reference-user, thus avoiding a subjective bias. In some cases, a voting scheme may be employed, to provide a consensus among reference-users. However, assuming that a reference-user does in fact have superior capabilities, the voting or statistical averaging may significantly diminish the advantage of having a reference-user with superior insight; such users may be capable of reacting outside of the statistical norms to great benefit, and therefore this advantage should not be squandered by requiring that the reference-user conform to statistical norms or models. Likewise, care should be employed when excluding outliers, since these may represent valuable opportunity. Whether to permit statistical deviation from the norm, or proceed by consensus, is a design decision in the system.

According to another aspect of the technology, a large data set may be processed to define a reduced data set based on reliability and coverage of the data space. The goal is not to place every available data point of the data set within the data space, but rather to define a filtered data set with maximum reliable coverage. Thus, portions of the data space densely populated with high reliability data generally have a higher threshold for inclusion of new data, while portions with lower reliability or lower density more readily accept new data. In this way, reliable statistical inferences can be efficiently drawn, using feasible analysis. Metrics and algorithms are provided for representing the relative veracity and usefulness of individual instances of information and the providing sources. The veracity of information is measured by the difference, if any, between which it disagrees with an overall “best estimate” calculated based on the preexisting data set. The usefulness of information is measured by the amount by which the instance of information decreases the amount of uncertainty. A reference-user may interact with this dataset to criteria regarding the density, veracity and usefulness criteria, influence data inclusion, and/or to cluster the data within the set. In general, correctness is determined by engineering techniques such as total quality management (TQM) and Truth Seeking (triangulation) principles in continuous monitoring. Data accuracy needs to be measured not only at individual data point level, but also when calculating derivative data points. This technology may be used in an asset database system to permit investment analysis, portfolio analysis, and peer analysis, for example.

Based on this reduced data set, peer groups of multivariate data are automatically determined using criteria relevant for human consideration, that is the data is projected into a low dimensional cognitive space. The reduced data set may be supplemented with an overlay of additional data (that is, similar data which is not in the reduced data set), which can then be subjected to the peer group analysis as well. The system is also appropriate for ad hoc queries, though common queries are pre-cached and thus results are available with low latency. The peer clustering algorithms, and the reduced data set may each be modified adaptively, providing dynamically updated analysis tools. The system preferably supports queries relating to the data and relationships of the data, as well as providing customizable dashboards representing common metrics of interest to the alternative investment community.

In order to automatically synthesize investment rating/grading of objects that represent investments, a distance function or transformation function is generated off the data set. As the data set changes, the distance function evolves such that all points with the same net risk or risk reward profile, map to the same cluster or point in the transformed space. The distance function is adaptive and “user evolvable”. This consists of a) a reference-user who trains the distance function b) a general group of users that continuously provide data and feedback on its results. The automated risk report for a particular asset is generated by finding all assets that have a similar net risk, i.e., are the same distance radius distance from the investment risk point. This cluster of points may then be rank ordered according to the return metric. The rating is then the “alpha”, or excess return over the average representation of the cluster of similar points.

According to one aspect, a mapping algorithm maps the multivariate discrete, continuous hybrid space representing the various factors that distinguish various risk reward profiles into a univariate normalized space such that it is now possible to apply asset allocation principles.

Intelligent Advertising

The value of an alternative asset (poorly valued because of an inefficient market) is the actually realized value at duration of the final exit for a party, as opposed to price, which is the transaction value attributed at the trade or transaction today. When we use this in the context of digital assets such as domain names, Google rankings, ad placement etc. all of which classify as alternatives because they are traded in an inefficient market, then the price is the price paid by the advertiser. If the search engine makes its advertising placement decision based on the highest advertising price only, over the long term this results in poorer placement of items and attrition of eyeballs, in effect reducing the value of the asset. Thus, understanding the difference between price and value, even directionally is critical. Accordingly, another aspect of the technology is to optimize advertisem*nt placement into a natural result (that is, not influenced by the advertising) by referring to the clustering of the data as well as the context, such that the advertising is appropriate, inoffensive, and likely to increase the overall value of the enterprise, based on both the short term revenues from advertising, and the long term reputation and future cash flows that may be influenced. For example, an inappropriately placed ad will generate advertising revenue, but may disincentivize the advertiser to place ads in the future. An appropriately placed ad, which is contextually appropriate and topically appropriate, is more likely to result in a consumer-advertiser transaction, and thus lead to higher future advertising revenues, even if the present value of the ad is not the highest possible option.

A reference-user in this context may be a user who transacts with an advertiser. By matching users with a reference-user, within the appropriate context, it is more likely that the users will also transact with that advertiser, as compared to users in a different context. The ads may therefore be clustered as artificial insertions into the data universe, and clustered accordingly. When a user's corresponding reference-user(s) and cluster(s) of interest are identified, the advertisem*nts within those clusters may then be considered for delivery to the user.

Location-Context Search

According to an embodiment of the technology, location may be used as a context to define a reference-user, and the reference-user profile is then exploited to produce a system response. Thus, rather than iteratively or implicitly determining a relevant context for a user, a location cue, such as GPS location, Internet service provider (ISP) location, IP mapping, inverse domain name lookup, ZIP code, user demographic profile, or the like. The location may this be the present location or a reference location.

The location context is actually determined by the respective users themselves both for the current and the reference location. A particular user has a particular set of location contexts, e.g., given an ambiguous location such as “School Street”, a first user may have the reference location context as “School Street in Providence R.I., USA” where the first user's relative lives versus a second user who may have the reference location context as “School Street in Belmont, Mass., USA” where the second user's child goes to school. Both reference locations are contextually relevant to the particular users, but different between different users.

Based on the context, e.g., location, a data entry or response may be selectively processed. Thus, a New Yorker may use language in a different way than a Londoner. In order to interpret the language, profiles of reference-users with similar location references, i.e., selected based on the context, are analyzed for query response correspondence. For example, the reference-user profiles may be used to perform word translations or equivalencies, rank search results, select types of results, and the like. As an example, the first user's reference location is also more relevant to other users/reference user in the first user's cluster.

Once the meaning of the input is determined with some reliability, the next step is determining a useful output. Note that the context for interpretation of the input may differ from the context for producing a meaningful output; that is, the relevant reference-users need not be the same. For example, the New Yorker in London might seek, through a speech recognition system on a smartphone, a club for entertainment. Upon recognizing both location cues, i.e., the origin of the user (which may be accessible from a telephone number or ID, user profile, cookie, speech accent, etc.) and the current location of the user, a set of reference-users may be selected. These may include other New Yorkers in London who visit clubs. However, the set of reference-users is not so limited. The reference-users may be selected more broadly based on preferences, affinities, chronologies, and the like, and may include both visitors to London and natives. Using location tracking and e-commerce technology, information about what day a respective reference-user went to the club, how long her or she spent, what they ordered, how much they tipped, etc., may all be available information. This type of information may be correlated with the user's past history, inside and out of London. Of course, to the extent that explicit ratings of clubs are available, these may also be exploited, but these explicit ratings tend to display bias and are not statistically normalized or validated. Note that the reliability of explicit ratings may improve dramatically when broken down by context, e.g., the reference-user(s) responsible for the rating. In general, using a large body of available information for prospective reference-users, a cluster analysis is performed which may rank or weight different users with respect to their probative value toward resolving the query. Depending on the system implementation, some aspects of the cluster analysis may be performed in advance, and thus only final stages computer on demand. Thus, the context for generating the system response may be determined, and that context used to determine the cluster in which the user “resides”, which then defines the reference-user(s) to be referenced in formulating the response. Alternately, an affinity with a reference user or user(s) is determined, e.g., with a collaborative filter, and that set of reference-users used to determine the context cluster. In either case, the response is then generated based on the context cluster, which is statistically linked to the reference-users associated with that cluster. The favorite clubs for the reference-users are then presented as a response to the query, and may be ranked according to weightings derived from the analysis.

It is noted that systems of the type described are typically subsidized by advertising. Therefore, once the meaning of the query is determined, e.g., the user is looking for a club, a set of ads for clubs, club goers, or the user abstract from his goal directed activity, may be presented. In general, a club would not wish to solicit a patron who would not have fun; the tab and tip will be low, and future referrals absent. Likewise, a targeted ad of this type may be relatively expensive, and thus there would be incentive for advertisers to present ads that will likely yield return on investment. The system itself has as a goal to deliver relevant advertising, since user mistrust in the output will lead to loss of usage and advertising revenues. Given the generally aligned incentives, therefore, the advertisers themselves may be provide useful and rich context information. That is, in contrast to normal users, who will often not spend time training a third party system, advertisers may be willing to spend considerable time defining their preferred customers, and providing useful information for those customers. In cases where there is an incentive to “cheat”, that is, game the system to achieve an unnatural result, feedback from actual users and a system-imposed punishment may be effective. Thus, if a user is “pushed” to go to a club they do not enjoy, the user may end up being a bad customer (low tab and tip), and may help redefine the cluster so that user for which he or she becomes a reference-user have reduced likelihood of becoming patrons. Since the system may be quite interactive and ubiquitous, feedback may be in nearly real-time. Of course, permitting advertisers to feed the system with information is merely optional, and therefore to the extent that some users seek to taint the system, the cluster analysis and context sensitivity may exclude other users from adverse impact.

Advertisers can target the most contextually relevant reference and current location to push particular content to a respective user.

Recommendation Engine

In another embodiment, a user seeks a recommendation from a recommendation engine. The recommendation engine contains identifications and profiles of users who have posted recommendations/ratings, as well as profiles for users and usage feedback for the system. A user seeking to use the engine is presented (at some time) with a set of questions or the system otherwise obtains data inputs defining the characteristics of the user. In this case, the user characteristics generally define the context which is used to interpret or modify the basic goal of the user, and therefore the reference-user(s) for the user, though the user may also define or modify the context at the time of use. Thus, for example, a user seeks to buy a point-and-shoot camera as a gift for a friend. In this case, there are at least four different contexts to be considered: the gift, the gift giver, the gift receiver, and the gifting occasion. The likelihood of finding a single reference-user appropriate for each of these contexts is low, so a synthetic reference-user may be created, i.e., information from multiple users and gifts processed and exploited. The issues for consideration are; what kinds of cameras have people similarly situated to the gift giver (the user, in this case) had good experiences giving? For the recipient, what kinds of cameras do similar recipients like to receive? Based on the occasion, some givers and recipients may be filtered. Price may or may not be considered an independent context, or a modifier to the other contexts. The various considerations are used in a cluster analysis, in which recommendations relevant to the contexts may be presented, with a ranking according to the distance function from the “cluster definition”. As discussed above, once the clustering is determined, advertisem*nts may be selected as appropriate for the cluster, to provide a subsidy for operation of the system, and also to provide relevant information for the user about available products.

Once again, the context is specific to the particular user and thus the right kind of camera for a first user to give a friend is not the same as the right kind of camera for a second user to give to a different friend; indeed, even if the friend is the same, the “right” kind of camera may differ between the two users. For example, if the first user is wealthier or other context differences.

One embodiment provides a decision support system, corresponding to the method shown in FIG. 2. A user input port receives user inputs, which define a user criterion or criteria, and also at least one user input tuning parameter. This parameter represents user tradeoff preferences for producing an output from a system 201. The output is may be in the form of tagged data, selected in dependence on the at least one user criterion, the at least one user input tuning parameter, and a distance function 202. A reference-user input is also provided which receives one or more reference-user input parameters representing a respective reference-user's analysis of the tagged data and the corresponding user inputs 203. The reference-user input is used to adapt the distance function in accordance, using the reference-user inputs as a feedback signal. The reference-user thus acts to optimize the distance function based on the user inputs and the output, and on at least one reference-user inference. This inference may be derived from at least one human user selected from a plurality of users based on an accuracy of selection according to an objective criterion 204. An information repository 205, such as a structured query language database, or so-called “No-SQL” database, configured to store the tagged data.

Another embodiment provides a decision support system, also generally corresponding to the method shown in FIG. 2, having a user input port configured to receive user inputs including at least one user criterion and at least one user input tuning parameter representing user tradeoff preferences 201. The user inputs are used to produce an output of tagged data in dependence on the at least one user criterion, the at least one user input tuning parameter, and a distance function 202. A reference-user agent is provided, which is configured to receive at least one reference-user input parameter representing the at least one reference-user's analysis of the tagged data and the corresponding user inputs 203. The reference-user agent selectively adapts the distance function in accordance with the user inputs as a feedback signal. That is, the user inputs are not necessarily directly used to provide feedback, but rather are filtered through the reference-user agent. The reference-user agent acts to optimize the distance function based on the user inputs and the output, and on at least one reference-user inference derived from at least one human user selected from a plurality of possible users based on an accuracy of selection according to an objective criterion 204. An information repository is provided, configured to store the tagged data 205.

A further embodiment provides a decision support method represented in FIG. 1, comprising receiving user inputs comprising at least one user criterion, and at least one user input tuning parameter representing user tradeoff preferences for producing an output 101; selectively producing an output of tagged data from a clustered database with dependence on at least one user criterion, the at least one user input tuning parameter, and a distance function 102; receiving at least one reference-user input parameter representing the at least one reference-user's analysis of the tagged data and the corresponding user inputs, to adapt the distance function in accordance with the reference-user inputs as a feedback signal 103; and clustering the database in dependence on at least the distance function 104, wherein the reference-user acts to optimize the distance function based on the user inputs and the output, and on at least one reference-user inference 105. The clustering may be automatically performed by a processor. The database may receive new data. The distance function may be applied to cluster the database including the new data before the at least one reference-user input parameter is received. The tagged data may comprise a valuation or rating. The distance function may be adaptive to new data. The reference-user inference may represent at least one of a valuation and a validation. The user input tuning parameter may comprise a dimensionless quantitative variable that impacts a plurality of hidden dimensions. The hidden dimensions may comprise at least one of completeness, timeliness, correctness, coverage, and confidence. The user input tuning parameter may balance completeness and correctness of the tagged data in the output.

Another embodiment provides an information access method, as shown in FIG. 3, comprising receiving a semantic user input comprising an indication of interest in information 301; determining a context of the user distinctly from the semantic user input comprising an indication of interest in information 302; producing an output of at least tagged data from a clustered database in dependence on at least the user input, the determined context, and a distance function 303; monitoring a user interaction with the output 304; and modifying the distance function in dependence on at least the monitored user interaction 305. The method may further comprise selecting at least one commercial advertisem*nt extrinsic to the tagged data from the clustered database for presentation to the user, in dependence on at least: at least one of the semantic user input and the output of tagged data, and the determined context 306. The selecting may be further dependent on the distance function. The monitoring may comprises monitoring a user interaction with the at least one commercial advertisem*nt, wherein the commercial advertisem*nt is selected in dependence on the distance function, and the distance function is modified based on the user interaction with a selected advertisem*nt 307. The method may further comprise reclustering the database in dependence on the modified distance function. The method may further comprise classifying a plurality of users, and distinguishing between difference classes of users with respect to the selection and modifying of respective ones of a plurality of distance functions 308. The method may further comprise determining at least one reference-user from a set of users, based on at least one fitness criterion, and selectively modifying the distance function dependent on a reference-user input in preference to a non-reference-user input 309. A user input is associated with a respective reference-user in dependence on the context 310.

Another embodiment provides an information processing method, as shown in FIG. 5, comprising: clustering a database comprising a plurality of information records according to semantic information contained therein, wherein information may be classified in a plurality of different clusters in dependence on a context, such that a common semantic query to the database yields different outputs over a range of contexts 501; producing an output identifying information records from the database in dependence on at least a user semantic input, and a distance function 502; receiving user feedback 503; and modifying at least one distance function in dependence on the user feedback 504. The method may further comprise determining a contextual ambiguity from the user semantic input, soliciting contextual ambiguity resolution information from the user, and producing a follow-up output identifying information records from the database in dependence on at least a user semantic input, the contextual ambiguity resolution information, and at least one distance function selected from a plurality of available distance functions in dependence on the contextual ambiguity resolution information 505. The method may further comprise selecting at least one commercial advertisem*nt extrinsic to the information records in the database for presentation to the user, in dependence on at least: the user semantic input, and the contextual ambiguity resolution information 506. The selecting may be further dependent on at least one distance function. The method may further comprise selecting at least one commercial advertisem*nt extrinsic to the information records in the database for presentation to the user, in dependence on at least: the user semantic input, and the distance function 507. The monitoring may comprise monitoring a user interaction with at least one commercial advertisem*nt presented to the user as part of the output, wherein the commercial advertisem*nt is selected in dependence on at least one distance function, and the at least one distance function is modified based on the user interaction with at least one selected advertisem*nt 508. The method may further comprise reclustering the database in dependence on the at least one modified distance function 509. The method may further comprise classifying a plurality of users, and distinguishing between difference classes of users with respect to the selection and modifying of respective ones of a plurality of distance functions 510. The method may further comprise assigning a reference-user status to at least one user within a set of users, based on at least one fitness criterion, and selectively weighting a user contribution to a modification of a respective distance function dependent on the reference-user status of the respective user 511. The reference-user status may be assigned with respect to a context, and a user input is associated with a respective distance function in dependence on the context 512.

Hardware Overview

FIG. 4 (see U.S. Pat. No. 7,702,660, issued to Chan, expressly incorporated herein by reference), shows a block diagram that illustrates a computer system 400 upon which an embodiment of the invention may be implemented. Computer system 400 includes a bus 402 or other communication mechanism for communicating information, and a processor 404 coupled with bus 402 for processing information. Computer system 400 also include a main memory 406, such as a random access memory (RAM) or other dynamic storage device, coupled to bus 402 for storing information and instructions to be executed by processor 404. Main memory 406 also may be used for storing temporary variables or other intermediate information during execution of instructions to be executed by processor 404. Computer system 400 further may also include a read only memory (ROM) 408 or other static storage device coupled to bus 402 for storing static information and instructions for processor 404. A storage device 410, such as a magnetic disk or optical disk, is provided and coupled to bus 402 for storing information and instructions.

Computer system 400 may be coupled via bus 402 to a display 412, such as a cathode ray tube (CRT), for displaying information to a computer user. An input device 414, including alphanumeric and other keys, is coupled to bus 402 for communicating information and command selections to processor 404. Another type of user input device is cursor control 416, such as a mouse, a trackball, or cursor direction keys for communicating direction information and command selections to processor 404 and for controlling cursor movement on display 412. This input device typically has two degrees of freedom in two axes, a first axis (e.g., x) and a second axis (e.g., y), that allows the device to specify positions in a plane.

The invention is related to the use of computer system 400 for implementing the techniques described herein. According to one embodiment of the invention, those techniques are performed by computer system 400 in response to processor 404 executing one or more sequences of one or more instructions contained in main memory 406. Such instructions may be read into main memory 406 from another machine-readable medium, such as storage device 410. Execution of the sequences of instructions contained in main memory 406 causes processor 404 to perform the process steps described herein. In alternative embodiments, hard-wired circuitry may be used in place of or in combination with software instructions to implement the invention. Thus, embodiments of the invention are not limited to any specific combination of hardware circuitry and software.

The term “machine-readable medium” as used herein refers to any medium that participates in providing data that causes a machine to operation in a specific fashion. In an embodiment implemented using computer system 400, various machine-readable media are involved, for example, in providing instructions to processor 404 for execution. Such a medium may take many forms, including but not limited to, non-volatile media, volatile media, and transmission media. Non-volatile media includes, for example, optical or magnetic disks, such as storage device 410. Volatile media includes dynamic memory, such as main memory 406. Transmission media includes coaxial cables, copper wire and fiber optics, including the wires that comprise bus 402. Transmission media can also take the form of acoustic or light waves, such as those generated during radio-wave and infra-red data communications. All such media must be tangible to enable the instructions carried by the media to be detected by a physical mechanism that reads the instructions into a machine. Non-transitory information is stored as instructions or control information.

Common forms of machine-readable media include, for example, a floppy disk, a flexible disk, hard disk, magnetic tape, or any other magnetic medium, a CD-ROM, any other optical medium, punchcards, papertape, any other physical medium with patterns of holes, a RAM, a PROM, and EPROM, a FLASH-EPROM, any other memory chip or cartridge, a carrier wave as described hereinafter, or any other medium from which a computer can read.

Various forms of machine-readable media may be involved in carrying one or more sequences of one or more instructions to processor 404 for execution. For example, the instructions may initially be carried on a magnetic disk of a remote computer. The remote computer can load the instructions into its dynamic memory and send the instructions over a telephone line using a modem. A modem local to computer system 400 can receive the data on the telephone line and use an infra-red transmitter to convert the data to an infra-red signal. An infra-red detector can receive the data carried in the infra-red signal and appropriate circuitry can place the data on bus 402. Bus 402 carries the data to main memory 406, from which processor 404 retrieves and executes the instructions. The instructions received by main memory 406 may optionally be stored on storage device 410 either before or after execution by processor 404.

Computer system 400 also includes a communication interface 418 coupled to bus 402. Communication interface 418 provides a two-way data communication coupling to a network link 420 that is connected to a local network 422. For example, communication interface 418 may be an Integrated Services Digital Network (ISDN) card or a modem to provide a data communication connection to a corresponding type of telephone line. As another example, communication interface 418 may be a local area network (LAN) card to provide a data communication connection to a compatible LAN. Wireless links may also be implemented. In any such implementation, communication interface 418 sends and receives electrical, electromagnetic or optical signals that carry digital data streams representing various types of information.

Network link 420 typically provides data communication through one or more networks to other data devices. For example, network link 420 may provide a connection through local network 422 to a host computer 424 or to data equipment operated by an Internet Service Provider (ISP) 426. ISP 426 in turn provides data communication services through the world wide packet data communication network now commonly referred to as the “Internet” 428. Local network 422 and Internet 428 both use electrical, electromagnetic or optical signals that carry digital data streams. The signals through the various networks and the signals on network link 420 and through communication interface 418, which carry the digital data to and from computer system 400, are exemplary forms of carrier waves transporting the information.

Computer system 400 can send messages and receive data, including program code, through the network(s), network link 420 and communication interface 418. In the Internet example, a server 430 might transmit a requested code for an application program through Internet 428, ISP 426, local network 422 and communication interface 418.

The received code may be executed by processor 404 as it is received, and/or stored in storage device 410, or other non-volatile storage for later execution.

In this description, several preferred embodiments were discussed. Persons skilled in the art will, undoubtedly, have other ideas as to how the systems and methods described herein may be used. It is understood that this broad invention is not limited to the embodiments discussed herein. Rather, the invention is limited only by the following claims.

The invention may be used as a method, system or apparatus, as programming codes for performing the stated functions and their equivalents on programmable machines, and the like. The aspects of the invention are intended to be separable, and may be implemented in combination, subcombination, and with various permutations of embodiments. Therefore, the various disclosure herein, including that which is represented by acknowledged prior art, may be combined, subcombined and permuted in accordance with the teachings hereof, without departing from the spirit and scope of the invention.