Data generalization for predictive models
Generalize the data through clustering technology and generate alternative feature sets, solving the problem of excessive data volume in the existing technology, and reducing the amount of data without reducing the performance of the prediction model, reducing the computing resources and storage requirements .
Patent Information
- Application Number
- CN202010783314.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2019-08-06
- Filing Date
- 2020-08-06
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2040-08-06
AI Technical Summary
The prior art is difficult to minimize the data provided to the prediction model without affecting the performance of the prediction model, especially in the face of large amounts of data and data privacy protection requirements, how to effectively reduce the amount of data stored and processed.
The data is generalized through clustering technology, alternative feature sets and generalized features are determined, and similar instances are grouped using algorithms such as K-means and hierarchical clustering to generate generalized feature sets, thereby reducing the amount of data stored and processed.
It realizes reducing the amount of data stored and processed without reducing the performance of the prediction model, reducing the computing resource and storage space requirements, while reducing the bandwidth and power consumption of data transmission.
Smart Images

Figure CN112347322B_ABST
Abstract
Description
Technical Field
[0001] The present disclosure relates generally to generalizing data, and more particularly to generalizing data for predictive models using clustering. Background Art
[0002] Data minimization can refer to the practice of limiting the collection of personal information to that which is directly relevant and needed to achieve a specific purpose. As companies and organizations begin to understand the power of data, and as data becomes more ubiquitous and easier to collect, analysts are faced with a flood of data. For a while, the impulse was to keep all of this data indefinitely. With the rapid adoption of smartphones, Internet of Things (IoT) devices, and the like, organizations are faced with more and more ways to collect more and more kinds of data, including, especially, private, personally identifiable data. Instead of keeping all data, data managers are now adopting a data minimization strategy to retain only the relevant and needed data. Summary of the invention
[0003] An exemplary embodiment of the disclosed subject matter is a method comprising: obtaining a labeled data set, wherein the labeled data set includes a plurality of instances and their labels, wherein each instance includes an evaluation of a feature set; based on the labels, determining a plurality of buckets, each of the plurality of buckets having an associated label, wherein each bucket groups instances from the labeled data set having labels within a similarity threshold with the label associated with the bucket; determining a plurality of clusters, wherein each cluster includes a plurality of instances included by the same bucket, wherein the determining the plurality of clusters is based on the evaluation of the feature set of the instances, thereby grouping similar instances into clusters; based on the plurality of clusters, determining an alternative feature set including a generalized feature set, wherein each generalized feature in the generalized feature set corresponds to a cluster in the plurality of clusters, wherein the generalized feature corresponding to a cluster indicates an instance that is a member of the corresponding cluster; obtaining a second instance; determining a generalized second instance, wherein the generalized second instance includes an evaluation of the alternative feature set of the second instance; and determining a label of the second instance based on the generalized second instance.
[0004] Another exemplary embodiment of the disclosed subject matter is a computer program product, comprising a non-transitory computer-readable storage medium storing program instructions, which, when read by a processor, causes the processor to perform: obtaining a labeled data set, wherein the labeled data set includes a plurality of instances and their labels, wherein each instance includes an evaluation of a feature set; based on the labels, determining a plurality of buckets, each of the plurality of buckets having an associated label, wherein each bucket groups instances from the labeled data set having labels within a similarity threshold with the label associated with the bucket; determining a plurality of clusters, wherein each cluster includes a plurality of instances included by the same bucket, wherein the determining of the plurality of clusters is based on the evaluation of the feature set of the instances, thereby grouping similar instances into clusters; based on the plurality of clusters, determining an alternative feature set including a generalized feature set, wherein each generalized feature in the generalized feature set corresponds to a cluster in the plurality of clusters, wherein the generalized feature corresponding to a cluster indicates an instance that is a member of the corresponding cluster; obtaining a second instance; determining a generalized second instance, wherein the generalized second instance includes an evaluation of the alternative feature set of the second instance; and determining a label of the second instance based on the generalized second instance.
[0005] Yet another exemplary embodiment of the disclosed subject matter is a computerized apparatus having a processor and coupled memory, the processor being adapted to perform the following steps: obtaining a labeled data set, wherein the labeled data set includes a plurality of instances and their labels, wherein each instance includes an evaluation of a feature set; based on the labels, determining a plurality of buckets, each of the plurality of buckets having an associated label, wherein each bucket groups instances from the labeled data set having labels within a similarity threshold with the label associated with the bucket; determining a plurality of clusters, wherein each cluster includes a plurality of instances included by the same bucket, wherein the determining the plurality of clusters is based on the evaluation of the feature set of the instances, thereby grouping similar instances into clusters; based on the plurality of clusters, determining an alternative feature set including a generalized feature set, wherein each generalized feature in the generalized feature set corresponds to a cluster in the plurality of clusters, wherein the generalized feature corresponding to a cluster indicates an instance that is a member of the corresponding cluster; obtaining a second instance; determining a generalized second instance, wherein the generalized second instance includes an evaluation of the alternative feature set of the second instance; and based on the generalized second instance, determining a label for the second instance. BRIEF DESCRIPTION OF THE DRAWINGS
[0006] The subject matter of the present disclosure will be more fully understood and appreciated through the following detailed description in conjunction with the accompanying drawings, in which corresponding or similar numbers or characters indicate corresponding or similar components. Unless otherwise indicated, the accompanying drawings provide exemplary embodiments or aspects of the present disclosure and do not limit the scope of the present disclosure. In the accompanying drawings:
[0007] Figure 1A and Figure 1B shows buckets and clustering according to some exemplary embodiments of the disclosed subject matter;
[0008] Figure 1C Decision diagrams illustrating some exemplary embodiments according to the disclosed subject matter;
[0009] Figure 1D A diagram showing clustering according to some exemplary embodiments of the disclosed subject matter;
[0010] Figure 2 a flowchart illustrating a method according to some exemplary embodiments of the disclosed subject matter; and
[0011] Figure 3 A block diagram showing an apparatus according to some exemplary embodiments of the disclosed subject matter is shown. DETAILED DESCRIPTION
[0012] One technical problem addressed by the disclosed subject matter is minimizing the data collected from users. In some exemplary embodiments, an entity collecting data may be subject to General Data Protection Regulation (GDPR) regulations. Thus, an entity may be required to limit data collection to data needed in relation to the purpose for which the data may be processed. Additionally or alternatively, GDPR and similar regulations may require that the collected private data be stored and protected in certain ways. Thus, reducing the amount of data stored may reduce the liability of an entity.
[0013] Another technical problem addressed by the disclosed subject matter is minimizing the data being collected for a decision model. In some cases, it may be desirable to minimize data, such as the number of features being collected and their respective granularity, while still being able to utilize a decision model to provide quality decisions, predictions, and the like. As an example, the same data record may be utilized by a predictive model configured to predict a heart attack and a predictive model configured to predict the number of offspring of a person. Different types of data minimization may be applied to each decision model, such as different types of feature generalization. As an example, in the case where an age feature is generalized to a range of 5 years, a first prediction model may have a performance metric of more than 90%, while in the case where the age feature is generalized to such a range, a second prediction model may have a performance metric of less than 80%. As another example, one model may be sensitive to ages over 50 years and their exact values, while another model may be indifferent and provide similar predictions for all values over 40 years. In some cases, a decision model may be a manual model that is used by an entity to make its decisions, such as, for example, a decision made by a banker to approve or deny a loan, a decision made by a banker regarding authorizing a maximum credit, a decision made by a clerk to provide a quote for a service to be supplied, and the like.
[0014] Yet another technical problem addressed by the disclosed subject matter is to minimize the data provided to a predictive model without affecting a performance metric of the predictive model. It may be desirable to minimize the data without causing the performance metric of the predictive model to drop below a threshold. The threshold may be an absolute threshold, such as 90%, 92%, etc. Additionally or alternatively, the threshold may be a relative threshold relative to the performance metric of the predictive model before data minimization. As an example, it may be desirable that after minimizing the data, the drop in the performance metric should not exceed 5%.
[0015] One technical solution is to determine a generalization of a feature set utilized by a decision model. In some exemplary embodiments, instead of determining the feature set, an alternative feature set may be determined. The alternative feature set may include a generalized feature set. The generalized feature may be a generalization of one or more features in the feature set. Generalizing the feature set may minimize data.
[0016] In some exemplary embodiments, a clustering process may be used to determine the alternative feature set. Generalized features may be associated with clusters determined during the clustering process.
[0017] Another technical solution is to perform a two-stage clustering process. The first stage can be based on the labels included by the labeled data set, and a plurality of buckets can be generated, each of which can be associated with a value of the label. The second stage can be based on an evaluation of the features of the instances included by the labeled data set.
[0018] In some exemplary embodiments, the second stage may be applied to the instances included in at least one bucket. Clustering algorithms such as K-means, hierarchical clustering, affine propagation, etc. may be used.
[0019] In some exemplary embodiments, a bucket may group instances from a labeled data set that have labels that are within a similarity threshold of the label associated with the bucket. As an example, a bucket may be associated with an age label of 30 years old. The similarity threshold may define a range of 2 years, such that instances with age labels between 29 and 31 years old may be grouped into a bucket. In some exemplary embodiments, the similarity threshold may be an equality threshold, an identity threshold, etc., such that a bucket groups instances with exactly the same labels.
[0020] Note that grouping instances into buckets may be based on a similarity threshold of labels, and determining clusters may be performed at each bucket based on an evaluation of similarity of features. Additionally or alternatively, clusters may group instances included by the same bucket. In some exemplary embodiments, each cluster may be associated with a bucket. Additionally or alternatively, there may be multiple determined clusters for a single bucket.
[0021] In some exemplary embodiments, a second instance may be obtained, and a label of the second instance may be determined. The second instance may be an unseen instance, such as an instance not included by a labeled data set, an instance for which a label is not available and for which a label needs to be determined, etc. In some exemplary embodiments, the second instance may be obtained without a known corresponding label. In some cases, generalizing the second instance may be based on an evaluation of an alternative feature set of the second instance. The evaluation of the alternative feature set may produce a corresponding generalized instance of the second instance. The generalized instance may indicate one or more clusters including the second instance, also referred to as the identified clusters. As an example, the second instance may be mapped to a single cluster (also referred to as a "member" of a single cluster). Additionally or alternatively, the second instance may be mapped to more than one cluster. In some exemplary embodiments, the label of the generalized instance may be determined based on the labels of one or more identified clusters. As an example, if the generalized instance indicates mapping to a single cluster, the label associated therewith, the average of the labels associated therewith, etc. may be determined as the label of the generalized instance. As another example, if a generalized instance indicates membership in two or more clusters, the label can be determined based on one or more labels of at least a portion of the identified clusters. In some exemplary embodiments, a membership metric can be calculated and used to determine the label. The membership metric can indicate which cluster the second instance is most associated with. As an example, it may be desirable to label an instance representing a person as a lawyer or a patent agent. The instance may have at least two features indicating a revenue channel, such as a feature indicating the person's income from litigation and a feature indicating the person's income from drafting patent applications. Several generalized features can be determined, each of which can correspond to a different cluster. Some clusters may include instances labeled as "lawyers", and some clusters may include instances labeled as "patent agents". The instances to be labeled can be analyzed and determined to be mapped to two different clusters, the first cluster being associated with the "lawyer" label and the second cluster being associated with the "patent agent" label. A membership metric can be defined to measure the membership of instances in each cluster. As an example, a membership function can be defined by measuring each revenue channel. In the case where a person's income from litigation is higher than the person's income from drafting (e.g., with or without normalization), it can be determined that the instance is more of a member of the attorney cluster, and the instance can be labeled as attorneys. On the other hand, in the case where a person's income from drafting is higher than the person's income from litigation, it can be determined that the instance is more of a member of the patent agent cluster than the attorney cluster, and the instance can be labeled as patent agent. It may be noted that the membership metric may return the same metric on both clusters (e.g., a person's income from litigation is equal to the person's income from drafting, or the two normalized incomes are equal).In this case, the instance is likely an outlier and the label can be randomly determined between the two options.
[0022] Additionally or alternatively, by measuring the distance from the centroid of each cluster, it can be determined whether the person is more representative of a patent agent or an attorney. In some cases, the generalization feature can include a distance measure from the centroid of each cluster. The label can be determined based on the label of the cluster to which the instance is closest (e.g., the identified cluster with the lowest value of the corresponding generalization feature). Additionally or alternatively, the label can be determined based on a weighted average or other calculation that takes into account a similarity measure of the instance to each identified cluster, as may be demonstrated by the value of the generalization feature.
[0023] In some exemplary embodiments, in order to determine the label of an instance, a specific instance may be determined. An instance may be generalized to a generalized instance in order to minimize data exposure. Based on the generalized instance, a specific instance may be determined. A specific instance may include an evaluation of the original feature, rather than an evaluation of an alternative feature set for representing the generalization. A specific instance may be provided to a decision model in order to determine the label of the instance. The decision model may be an automatic model, a predictive model, a manual decision process performed by a person, etc. In some cases, a specific instance may be a statistical representation of an observed instance, and the observed instance is a member of the identified cluster. The observed instance may be an instance previously observed during the practice of the disclosed subject matter, such as included by a labeled data set, provided to a decision model as part of a training data set, etc. The statistical representation of the observed instance may be an average value of the observed instance, a mode of the observed instance, a mean of the observed instance, etc. Additionally or alternatively, a cluster including an instance may have a centroid, and a cluster may include a centroid. An example of a cluster including its centroid may be a convex cluster. In this case, the centroid may be a specific value. Additionally or alternatively, in the case where an instance is included by multiple identified clusters, a calculation based on the observed instances, the centroid of each of them, etc. can be performed to generate a specific instance. As another example, the generated specific instance can be determined to be included by all clusters that include the original instance, such as can be generated based on the intersection of all identified clusters. The specific instance can be generated as the centroid of the intersecting clusters based on the instances observed in the cluster of interest, etc.
[0024] In some exemplary embodiments, when the number of observed instances is below a threshold, the specific instance may be the centroid. Once the number of observed instances exceeds the threshold, the specific instance may be a statistical representation of the observed values.
[0025] In some exemplary embodiments, a predictive model may be used to determine a labeled data set utilized by the disclosed subject matter. In some exemplary embodiments, an unlabeled data set may be obtained. The predictive model may be applied to the unlabeled data set. Using the predictive model, a predicted label for each instance in the unlabeled data set may be determined, thereby generating a labeled data set. This may be in contrast to conventional labeled data sets, which are typically labeled with accurate and correct labels rather than with predicted labels.
[0026] In some exemplary embodiments, an instance can be obtained and generalized on the edge device. The generalization of the instance can be transmitted to a server, and a label of the second instance can be determined on the server.
[0027] One technical effect of utilizing the disclosed subject matter is that the data that a decision model may need to determine a label is minimized. Instead of leaking complete information, a minimized representation thereof may be leaked, thereby reducing private information stored and processed.
[0028] Another technical effect of utilizing the disclosed subject matter is that the computing resources required to predict the label of an instance are reduced. In some exemplary embodiments, instead, if a prediction model is applied, relatively non-capacity intensive calculations can be applied based on the identified clusters. In some cases, the prediction model can be an ANN (artificial neural network) model (e.g., CNN (convolutional neural network), RNN (recurrent neural network), DNN (deep neural network), etc.), a non-ANN model (e.g., decision tree, SVM (support vector machine), etc.). In some exemplary embodiments, the disclosed subject matter can provide predictions using a reduced amount of computing resources compared to the resources required to apply the prediction model itself.
[0029] Another technical effect of utilizing the disclosed subject matter is that the storage space required to store data, the bandwidth required to transmit data, and the power required to transmit data are reduced. In some exemplary embodiments, data instances can be saved for future use, such as domain-specific use, quality assurance, training of other models, retraining of predictive models, and the like. Additionally or alternatively, it may be desirable to transmit the data instances to a remote server. As an example, data instances can be obtained from an IoT device. IoT devices can measure temperature, humidity, light, and the like once per second. Instead of saving exact values, re-encoding can be applied to represent more generalized data using a reduced number of bits. In some cases, generalization can be viewed as a form of lossy compression that loses insignificant information about the decision model used by the server that collects data from the IoT device.
[0030] The disclosed subject matter may provide one or more technical improvements over any prior art and any technology that has previously become conventional or traditional in the art.
[0031] In view of this disclosure, additional technical problems, solutions, and effects may be apparent to those of ordinary skill in the art.
[0032] Reference now Figure 1A and Figure 1B , Figure 1A and Figure 1B Buckets and clusters are shown according to some exemplary embodiments of the disclosed subject matter.
[0033] Figure 1A Instances are shown, such as instance 103. Each instance includes a value for each feature in the feature set. For example, instance 103 has an evaluation {age=30, gender=M, hair color=brown}. Each instance can be associated with a label, such as true or false for a Boolean attribute, a salary attribute, etc.
[0034] Instances may be assigned to buckets, such as buckets 101, 102. Each bucket includes instances with the same label or with similar labels, such as labels with a similarity metric that does not exceed a threshold for the labels of the bucket. For example, bucket 101 may include all instances with a true label, while bucket 102 may include all instances with a "false" label. As another example, bucket 101 may include all instances with a salary of 30K-40K, while bucket 102 may include all instances with a salary of 40K-50K or more.
[0035] Figure 1B The instances included by bucket 101 are shown clustered into three clusters: clusters 111, 112, and 113. Clustering may group instances having similar evaluations of a set of features. For example, cluster 111 may include instances having similar hair color and similar ages between 30 and 35 years old. In some exemplary embodiments, a person between the ages of 30 and 35 years old with brown or black hair may not be required to disclose their exact age or hair color, while still allowing the disclosed subject to provide their relatively accurate label. Since they are members of the cluster in bucket 101, the label may be determined based on the label of bucket 101 without requiring precise information about age and hair color. Additionally or alternatively, specific instances of potential alternatives that are similarly located within cluster 111 may be generated, and such specific instances may be used to determine the label of the actual instance, for example, by applying a decision model to the specific instances of the potential alternatives. Note that in Figure 1B In the example shown in , generalization to other age ranges may not be possible based on current data, current clustering, etc. As an example, people between the ages of 40 and 60 may have to reveal their exact age.
[0036] In some exemplary embodiments, the larger the data set and the fewer outliers present in the data, the more likely it is that generalization can be made using clustering.
[0037] In some exemplary embodiments, generalization may be provided by indicating to which cluster an instance belongs. Additionally or alternatively, the generalization may include, for each cluster, an indication of whether the instance belongs to the cluster, and a membership metric associated therewith.
[0038] In some exemplary embodiments, the generated clusters (and therefore the alternative feature sets) may not cover the entire domain, so it is possible that a data point may be received later that does not map to any cluster. In this case, it is possible to either not generalize such data point, or map it to the closest cluster, thus achieving generalization but possibly losing some accuracy.
[0039] Reference now Figure 1C , Figure 1C An illustration showing a decision diagram according to some exemplary embodiments of the disclosed subject matter.
[0040] It can be based on buckets and clusters (e.g. Figure 1A and Figure 1B The decision graph 120 can be shown in FIG. Figure 1A and Figure 1B An alternative feature set corresponding to the feature set in . Figure 1A and Figure 1B An example with 3 features is shown. For simplicity, decision diagram 120 shows 2 alternative features: age and gender. Decision diagram 120 generalizes the age feature. The age feature is represented using the following four subdomains: (20-25), (25-30), (30-35), (35-40). Each region can correspond to a different label. Instances in region 122 or region 132 can have a first label, instances in region 124, region 128, or region 134 can have a second label, and instances in region 126 can have a third label. As an example, an instance may have an evaluation of "female" for the gender feature, and a value of 23 years old as an evaluation for the age feature. The instance can be located in region 122 and have a first label.
[0041] In some exemplary embodiments, for some features, other features may be further generalized, resulting in conditional generalizations. As an example, as can be seen in decision diagram 120, in the case where the instance has a value of "female" for gender, the generalized age feature may have the following two subdomains instead of four subdomains: (20-30) and (30-40). In some exemplary embodiments, the disclosed subject matter may be used in order to minimize the data obtained from a user. In the case where the user is female, it may be sufficient to know whether her age is between 20 and 30 years old or between 30 and 40 years old.
[0042] In some exemplary embodiments, the decision graph 120 may be generated based on clustering. In some exemplary embodiments, the decision graph 120 may include regions, each region corresponding to a different cluster. In such a case, the regions may have potentially irregular shapes, may overlap, portions of the space may not be associated with any region, etc. In some exemplary embodiments, the decision graph 120 may be generated as an abstraction of clustering, for example, by defining lines that separate clusters associated with different labels. In some exemplary embodiments, the space defined by the decision graph 120 may be divided into regions using straight lines, such that each region substantially includes clusters associated with the same label. In some cases, lines parallel to the axis may be utilized in order to provide a clear division of the values of the axis into subdomains.
[0043] Reference now Figure 1D , Figure 1D A diagram showing clustering according to some exemplary embodiments of the disclosed subject matter.
[0044] In some exemplary embodiments, a two-dimensional feature space may be shown to visually show clusters. A cluster may be a shape in a feature space. In a two-dimensional space, the shape may be a two-dimensional shape, which may be symmetrical or asymmetrical. Cluster 152 illustrates a cluster in the form of a circle. The centroid 152c, which is the center of cluster 152, is inside cluster 152. When calculating the label of a generalized instance included by cluster 152, centroid 152c may be used as a specific instance. Additionally or alternatively, cluster 156 may not be a convex grouping. The centroid 156c of cluster 156 is outside cluster 156. In some exemplary embodiments, when a specific instance of cluster 156 is generated, centroid 156c may not be used as is. In some cases, an alternative observed instance may be selected. Additionally or alternatively, the observed instance closest to centroid 156c may be used. Additionally or alternatively, it may be used only when the centroid of a cluster is observed. Otherwise, an alternative observed instance that is within the cluster and closest to the cluster may be used.
[0045] Reference now Figure 2 , Figure 2 A flow chart illustrating a method according to some exemplary embodiments of the disclosed subject matter.
[0046] At step 200, a labeled data set may be obtained. The labeled data set may include instances and their labels. In some exemplary embodiments, the labels of the labeled data set may be obtained by applying a decision model to a data set of unlabeled instances. In some exemplary embodiments, the decision model may be a prediction model configured to determine a label of an instance, thereby providing a predicted label that may or may not be accurate. Additionally or alternatively, the labeled data set may be a labeled data set collected from any other source or using any other method.
[0047] In step 210, a plurality of buckets may be determined. Each bucket may be associated with a label. In some exemplary embodiments, each bucket may group instances comprised by a labeled data set having the same label. Additionally or alternatively, a bucket may include instances having similar but not necessarily identical labels. In some exemplary embodiments, the grouping of instances may be based on a similarity threshold. The threshold may define how similar the label of the instance and the associated label of the bucket should be. In some exemplary embodiments, the similarity threshold may be an equality threshold so that only instances having exactly the same label associated with the bucket may be grouped into the bucket. In some exemplary embodiments, a similarity measure may be calculated based on the label associated with the bucket and based on the label of each instance. The similarity measure may be compared to the similarity threshold to determine whether the instance is a member of the bucket.
[0048] In step 220, multiple clusters may be determined. Multiple clusters may be determined for each bucket. In some exemplary embodiments, multiple clusters may be determined separately and independently for each bucket. In some exemplary embodiments, a single instance may be a member of more than one cluster. Additionally or alternatively, there may or may not be overlap between these clusters. In some exemplary embodiments, each cluster may be included by exactly one bucket. Each cluster may include similar instances. The similarity of instances grouped into clusters may be based on an evaluation of a feature set of each instance. In some exemplary embodiments, any clustering technique may be used to perform the determination of clusters, such as, but not limited to, k-means clustering, centroid-based clustering, distribution-based clustering, density-based spatial clustering of applications with noise (DBSCAN), balanced iterative reduction and clustering using stratification (BIRCH), and the like.
[0049] At step 230, an alternative feature set may be determined. In some exemplary embodiments, the alternative feature set may be a generalization of the feature set. In some exemplary embodiments, the alternative feature set may include a portion of a feature, one or more generalized features, and the like. Additionally or alternatively, the alternative feature set may include generalized features. Generalized features may correspond to features included by the feature set. Features may have a range of possible values. Generalized features may have a reduced domain, including a smaller number of possible values, each of which corresponds to a subdomain in the domain. In some exemplary embodiments, generalized features may be determined based on clusters such as the clusters determined at step 220. In some exemplary embodiments, generalized features may be features that indicate membership of an instance to a cluster. In some exemplary embodiments, generalized features may indicate membership and membership metrics. For example, instead of a feature set that includes 50 different features, an alternative set may include 10 alternative features, each of which indicates membership of an instance to a different cluster. It may be noted that in some cases, there may be more alternative features than the original features. In some exemplary embodiments, using an alternative feature set may provide a reduced leakage of a user's private information. Such reduced leakage may be applicable even if the number of alternative features is greater than the number of original features.
[0050] At step 240, a second instance may be obtained. The second instance may be an unlabeled instance. In some exemplary embodiments, it may be desirable to determine a label for the second instance. In some exemplary embodiments, the second instance may be an instance that has not been previously used by the disclosed subject matter, such as not included in the labeled dataset of step 200. Additionally or alternatively, the second instance may be an instance that has not been used to train a predictive model used in accordance with the disclosed subject matter. In some exemplary embodiments, the second instance may be mapped to one or more identified clusters, for example, based on an evaluation of a set of features defined in the second instance.
[0051] At step 250, a generalized second instance may be determined. The generalized second instance may be a generalization of the second instance. In some exemplary embodiments, the generalized second instance may include an evaluation of an alternative feature set. In some exemplary embodiments, the generalized second instance may be determined by mapping an evaluation of features of the second instance to an evaluation of the alternative feature set. Additionally or alternatively, the generalized second instance may be determined based on membership of the second instance to a different cluster determined at step 220.
[0052] At step 260, a label for the generalized second instance may be determined. The label determined for the generalized second instance may be used as the label for the second instance. In some exemplary embodiments, determining the label for the generalized second instance may include performing steps 262 and 264. Additionally or alternatively, determining the label for the generalized second instance may include performing step 266.
[0053] In step 262, a specific instance may be generated. The specific instance may be generated based on the generalized second instance. The specific instance may include an evaluation of features including a feature set consistent with the generalized second instance. In some exemplary embodiments, the generalized second instance may indicate one or more clusters including the second instance, for example, via the value of the alternative feature set. The specific instance may be generated to provide an instance that is a member of one or more clusters including the second instance. The specific instance may potentially be different from the second instance. In some exemplary embodiments, the same generalization, the generalized second instance, may be utilized for both the specific instance and the second instance.
[0054] In some exemplary embodiments, the specific instance may be the centroid of the cluster including the second instance. Additionally or alternatively, the centroid may be utilized after verifying that the centroid is also a member of the cluster. Additionally or alternatively, in order to avoid utilizing false and unrealistic values, the observed instance may be utilized. The observed instance may be an instance observed during the application of the disclosed subject matter, such as an instance in a labeled data set, an obtained instance to be labeled (e.g., the second instance of step 240), etc. In some exemplary embodiments, the observed instance may be selected and used as a specific instance. The selection may be based on the relationship between the observed instance and the centroid or another statistical representation of all observed instances. For example, the observed instance closest to the centroid and included in the cluster may be selected and utilized. Additionally or alternatively, in the case where the generalized second instance indicates membership to more than a single cluster, a specific instance may be selected from the observed instances that are members of the same cluster set as the second instance. Additionally or alternatively, the top cluster may be determined based on a membership metric, and a specific instance in the top cluster may be generated.
[0055] At step 264, the model may be applied to the specific instance to predict a label for the second instance. In some exemplary embodiments, the model may be a decision model used to label the labeled data set of step 200. Additionally or alternatively, the decision model may be an automated model, a manual model, or the like, for example, implemented using machine learning techniques.
[0056] At step 266, labels may be determined based on a decision graph, such as decision graph 120. Each region in the decision graph may correspond to a label. By determining evaluations for alternative feature sets, generalized instances may be mapped to regions, and labels may be determined based on the regions.
[0057] Additionally or alternatively, the label may be determined based on the label of the cluster that includes the second instance. Additionally or alternatively, in the case where there are multiple clusters that include the second instance, their labels may be utilized, for example, by calculating their weighted average. The weights of the weighted average may be based on a measure of the membership of the second instance to different clusters, as may be recorded in the generalized second instance.
[0058] Reference now Figure 3 , Figure 3 A block diagram of an apparatus according to some exemplary embodiments of the disclosed subject matter is shown.
[0059] In some exemplary embodiments, the device 300 may include one or more processors 302. The processor 302 may be a central processing unit (CPU), a microprocessor, an electronic circuit, an integrated circuit (IC), etc. The processor 302 may be used to perform calculations required by the device 300 or any of its subcomponents.
[0060] In some exemplary embodiments of the disclosed subject matter, apparatus 300 may include an input / output (I / O) module 305. I / O module 305 may be used to provide output to a user and receive input from a user, such as, for example, obtaining a generalization instance, providing a prediction, etc. In some exemplary embodiments, I / O module 305 may be configured to obtain a prediction model, obtain a data set, obtain a labeled data set, etc. Additionally or alternatively, I / O module 305 may be configured to send a generalization instance.
[0061] In some exemplary embodiments, the apparatus 300 may include a memory unit 307. The memory unit 307 may be a hard drive, a flash disk, a random access memory (RAM), a memory chip, etc. In some exemplary embodiments, the memory unit 307 may store program code operable to cause the processor 302 to perform actions associated with any subcomponent of the apparatus 300. In some exemplary embodiments, the memory unit 307 may store data sets, metric performance results, etc. Additionally or alternatively, the memory unit 307 may store prediction models, buckets, clusters, etc.
[0062] Memory 307 may include one or more components described in detail below, implemented as executable files, libraries, static libraries, functions, or any other executable components.
[0063] In some exemplary embodiments, the bucket determiner 310 may be configured to determine a bucket for the labeled data set based on the label of the labeled data set. A bucket may be a data structure that holds instances, references to instances, etc. Each bucket may hold instances with the same label or instances with similar labels.
[0064] In some exemplary embodiments, the cluster determiner 320 may be configured to determine clusters of instances within a bucket. In some exemplary embodiments, the cluster determiner 320 may use any clustering technique to determine clusters, such as but not limited to k-means clustering, centroid-based clustering, distribution-based clustering, density-based spatial clustering of applications with noise (DBSCAN), balanced iterative reduction and clustering using hierarchical (BIRCH), etc. The cluster determiner 320 may utilize any clustering algorithm, such as but not limited to K-means, hierarchical clustering, affine propagation, etc. In some exemplary embodiments, the cluster determiner 320 may independently determine the clusters in each bucket determined by the bucket determiner 310.
[0065] In some exemplary embodiments, the alternative feature determiner 330 may be configured to determine an alternative feature set. The alternative feature set may be a generalization of the feature set. The alternative feature set may include generalized features having a generalized domain corresponding to a feature having a domain. Each generalized feature included by the alternative feature set may correspond to a cluster.
[0066] In some exemplary embodiments, the generalized instance determiner 340 may be configured to determine a generalized instance of a specific instance. In some exemplary embodiments, the generalized instance may have an evaluation of an alternative feature set. The generalized instance determiner 340 may receive an instance with an evaluation of features. Based on the evaluation and based on the alternative feature set, the generalized instance determiner 340 may output a generalized instance.
[0067] In some exemplary embodiments, the specific instance determiner 350 may be configured to determine the specific instance based on the generalized instance. In some exemplary embodiments, the generalized instance may indicate that it represents an instance that is a member in a cluster. The specific instance determiner 350 may determine another member of the cluster as a specific instance. In some exemplary embodiments, the specific instance may be determined as a statistical representation, such as a mode, mean, etc. of the instances included by the cluster, the centroid of the cluster, etc. Additionally or alternatively, the specific instance determiner 350 may be selected between the observed instances included by the cluster. Additionally or alternatively, in the case where the generalized instance indicates multiple clusters, the top-most cluster may be selected and the specific instance may be generated based on the top-most cluster. Additionally or alternatively, the specific instance may be generated based on multiple clusters, such as by generating an instance that is a member of all multiple clusters.
[0068] In some exemplary embodiments, the label determiner 360 may be configured to determine a label for a specific instance. The label determiner 360 may determine the label for a specific instance, for example, by applying a decision model. In some exemplary embodiments, determining the label may be performed on a server. The computerized edge device may be configured to obtain an instance and determine a generalized instance of the instance. The computerized edge device may transmit the generalized instance, the specific instance, etc. to the server so that the server determines the label of the instance based on the generalized instance. In such an embodiment, minimized data is transmitted to the server, thereby reducing the exposure of potential private information stored locally on the edge device.
[0069] In some exemplary embodiments, the server may determine the label by utilizing a prediction model. The prediction model may be trained based on the labeled data set. Additionally or alternatively, the server may transmit the prediction model to the edge device. In some exemplary embodiments, the edge device may be configured to locally determine the label of the instance based on the second instance without transmitting any data.
[0070] In some exemplary embodiments, exposing end users to mapping functions determined based on a decision model, calculation scopes determined based on a decision model, clustering determined based on a decision model, etc. may disclose some information about the decision model itself, about its training data, etc. In some cases, it may be desirable to prevent end users, client devices, etc. from gaining access to such information about the decision model, which may be a trade secret.
[0071] In some exemplary embodiments, function encryption can make it possible to encrypt the input to a function and calculate the result of the function by encrypting the data, while the party performing the calculation can only decrypt the result of the calculation. In some exemplary embodiments, the function encryption scheme for a given function f includes the following four algorithms:
[0072] (pk, msk)←setup(1 λ ):Create public key pk and master key msk.
[0073] ·sk←Keygen(m s k, f): Use the master key to generate a new key sk for function f.
[0074] c←Enc(pk, x): Encrypt message x using the public key.
[0075] y←Dec(sk, c ): Use the key to calculate y = f(x), where x is the encrypted value of c.
[0076] In some exemplary embodiments, a function encryption scheme may be used to calculate a function f, such as a function that maps an original data point to a corresponding cluster or range. In some exemplary embodiments, an edge device (e.g., a client terminal used by a user) may obtain data (d) from a user. The data (d) may be encrypted using a public key (pk). The edge device may transmit the encrypted data to a server, thereby preventing the server from accessing the data itself. The server may use a key (sk) to decrypt the result of applying the function f to the encrypted data (f(d)) to receive an associated cluster or feature range associated with the user. In some exemplary embodiments, using function encryption, client devices and users may not be able to access information that enables reverse engineering or understanding of a decision model (or a training data set used to train a decision model), while still protecting their privacy and ensuring that the actual data point (d) is not disclosed to the server.
[0077] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.
[0078] A computer-readable storage medium may be a tangible device that can hold and store instructions used by an instruction execution device. A computer-readable storage medium may be, for example, but not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples of computer-readable storage media (a non-exhaustive list) include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a raised structure in a groove on which instructions are stored, and any suitable combination of the foregoing. As used herein, a computer-readable storage medium is not to be interpreted as a transient signal per se, such as a radio wave or other freely propagating electromagnetic wave, an electromagnetic wave propagating through a waveguide or other transmission medium (e.g., a light pulse through a fiber optic cable), or an electrical signal transmitted through a wire.
[0079] The computer microscopic program instructions described herein can be downloaded from the computer microscopic storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as a serial, redirect, wide area network and / or wireless network. The network can include copper transmission cables, optical fiber transmission, wireless transmission, routers, firewalls, switches, gateway computers and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer replacement program instructions from the network and copies the computer fine program instructions for storage in the computer subdivided storage medium in each computing / processing device.
[0080] The computer program instructions for performing the operation of the present invention may be assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages such as Java, Smalltalk, C++, etc., and traditional procedural programming languages such as "C" language or similar programming languages. Computer-readable program instructions may be executed entirely on a user's computer, partially on a user's computer, as an independent software package, partially on a user's computer, partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect via the Internet). In some embodiments, an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA), or a programmable logic array (PLA), may be personalized by utilizing the state information of the computer-readable program instructions, and the electronic circuit may execute the computer-readable program instructions, thereby realizing various aspects of the present invention.
[0081] Various aspects of the present invention are described herein with reference to the flow charts and / or block diagrams of the methods, devices (systems) and computer program products according to embodiments of the present invention. It should be understood that each box of the flow chart and / or block diagram and the combination of each box in the flow chart and / or block diagram can be implemented by computer-readable program instructions.
[0082] These computer-readable program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processor of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.
[0083] Computer precise program instructions may also be loaded onto a computer, other data processing apparatus, or other device to perform various operating steps on the computer, other data processing apparatus, or other device to produce a computer-implemented process, whereby the instructions executed on the computer, other asynchronous data processing apparatus, or other device implement the functions / actions specified in one or more boxes in the flowchart and / or prompts.
[0084] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple embodiments of the present invention. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of the module, program segment or instruction includes one or more executable instructions for realizing the specified logical function. In some alternative implementations, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be implemented with a dedicated hardware-based system that performs the specified function or action, or can be implemented with a combination of special hardware and computer instructions.
[0085] The terms used herein are for the purpose of describing specific embodiments only and are not intended to limit the present invention. As used herein, the singular forms "a", "an" and "the" are also intended to include the plural forms, unless the context clearly indicates otherwise. It will be further understood that the terms "comprise" and / or "comprising" when used in this specification specify the presence of the features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or groups thereof.
[0086] All means or steps in the following claims plus corresponding structures, materials, actions, and equivalents of functional elements are intended to include any structure, material, or action for performing a function in combination with other claimed elements specifically claimed. The description of the present invention has been given for the purpose of illustration and description, but is not intended to be exhaustive or to limit the invention to the disclosed form. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the invention. The embodiments are selected and described in order to best explain the principles and practical applications of the invention, and to enable other persons of ordinary skill in the art to understand the various embodiments of the invention, which have various modifications suitable for the specific purposes contemplated.
Claims
1. A method comprising: Obtaining a labeled dataset, wherein the labeled dataset is related to age, the labeled dataset includes a plurality of instances and their labels, the labels including an age label, wherein each instance includes an evaluation of a feature set; Based on the labels, determining a plurality of buckets, each bucket of the plurality of buckets having an associated label, wherein each bucket groups instances from the labeled data set having labels within a similarity threshold to the label associated with the bucket; determining a plurality of clusters, wherein each cluster includes a plurality of instances included by a same bucket, wherein the determining the plurality of clusters is based on evaluating a set of features of the instances, thereby grouping similar instances into clusters; determining, based on the plurality of clusters, an alternative feature set comprising a set of generalized features, wherein each generalized feature in the set of generalized features corresponds to a cluster in the plurality of clusters, wherein the generalized feature corresponding to a cluster indicates an instance that is a member of the corresponding cluster; Get a second instance; determining a second instance of the generalization, wherein the second instance of the generalization includes evaluating an alternative feature set for the second instance; as well as Based on the generalized second instance, a label for the second instance is determined. 2 . The method of claim 1 , wherein the similarity threshold is an equality threshold, whereby each bucket groups instances from the labeled data set having an associated label for each bucket.
3. The method of claim 1 , wherein determining the label of the second instance comprises: determining an identified cluster based on an evaluation of the alternative feature set of the generalized second instance, wherein the second instance is mapped to the identified cluster; as well as Wherein the label of the second instance is determined based on the label associated with the bucket associated with the identified cluster.
4. The method of claim 3, wherein the determining the identified cluster comprises determining a plurality of identified clusters, wherein the label of the second instance is determined based on one or more labels of one or more buckets associated with the plurality of identified buckets.
5. The method of claim 1 , wherein determining the label of the second instance comprises: determining a specific second instance based on the generalized second instance; as well as The specific second instance is provided to the prediction model, wherein the prediction model is configured to provide a second label for the specific second instance. The method of claim 5 , wherein the specific second instance is different from the second instance.
7. The method of claim 5, wherein the specific second instance is a statistical representation of the observed instances comprised by the cluster.
8. The method of claim 5, wherein a cluster comprises a centroid, wherein the centroid is inside the cluster, and wherein the specific second instance is the centroid of the cluster.
9. The method according to claim 1, wherein obtaining the second instance and determining the generalized second instance are performed on the edge device; The method further comprises: transmitting, by the edge device, the generalized second instance to the server; as well as The determining of the label of the second instance based on the generalized second instance is performed on the server.
10. The method according to claim 1, wherein obtaining the labeled data set comprises: Get multiple instances; as well as Applying a prediction model to each instance of the plurality of instances to determine its label, wherein the prediction model is configured to determine the label of the instance.
11. A non-transitory computer-readable storage medium storing program instructions, which, when read by a processor, cause the processor to execute the method according to any one of claims 1 to 10.
12. A computerized device having a processor and coupled memory, the processor being adapted to perform the steps of the method according to any one of claims 1-10.
13. A system comprising modules respectively configured to perform the steps of the method according to any one of claims 1-10.
