Patent Application
20140317034
This disclosure generally relates to data classification. More particularly, and without limitation, this disclosure relates to predicting classification labels for a data set based on a smaller training data set.
DESCRIPTION OF THE RELATED ART
There are a variety of situations in which it would be useful to be able to efficiently predict class assignment for data based on limited training data resources. Typical approaches include using a training data set of measurements that has class labels assigned to the training set measurements. An additional set of class labels are generated for additional measurements with the goal of minimizing prediction error for additional measurement data.
In some cases the input data is given in the form of measurements X1, . . . ,XN, where each X is a vector of d measurements. A small subset of the input data is size L (L<k≦N) and may have attached class assignment labels: Y1, . . . ,YL, such that Yi's are binary integers (e.g., {−1,1}). This set is referred to as the ‘training set’. The task is then to augment the pairs (X1,Y1), . . . ,(XL,YL) with an additional minimal set of pairs by actively selecting and labeling them from XL+1, . . . , Xk, such that the prediction error (or risk) of a classifier output—Y′k+1, . . . ,Y′N using the new labeled training set is minimized for a given test set Xk+1, . . . ,XN.
The existing solutions belong to a class of learning algorithms defined in the literature as ‘active’ classifiers, as they actively construct a training set from which they can learn and predict the class labels of a given test set. Given limited labeling resources, the active classifier algorithm is configured to obtain an optimal label assignment for the test set while querying as few as possible training set members.
One shortcoming of existing solutions is that they are not efficient at handling large data sets. The time and computational resources required to obtain results in some circumstances detracts from the potential value of the results. For example, an existing solution may take over a month of running time using 20 CPUs for relatively small size data sets. Another feature of existing solutions is that they use different criteria in classification and query selection, which can yield classification accuracy that is less than desired.
An illustrative data classifier device includes data storage and at least one processor configured to predict classification labels for data. The processor is configured to determine a relationship between the data and training data with associated training classification labels. The processor is also configured to assign a weighted version of at least one of the training classification labels to at least one member of the data based on the determined relationship.
An illustrative method of classifying data includes using at least one processor for predicting classification labels for data by determining a relationship between the data and training data with associated training classification labels. A weighted version of at least one of the training classification labels is assigned to at least one member of the data based on the determined relationship.
Various embodiments and their features will become apparent to those skilled in the art from the following detailed description of at least one example embodiment. The drawings that accompany the detailed description can be briefly described as follows.
Some implementations will include pool data 32 while others will not. In situations that include pool data, the pool data 32 is that which can be used by the query engine 28 to identify an addition to the training data 30 (in a manner described below). In situations that do not include pool data 32, the query engine 28 considers the test data 34 to identify an addition to the training data 30. For discussion purposes, the following description will refer primarily to the test data 34 and that should be understood in a generic sense to include the pool data 32 if pool data were included unless the context clearly indicates otherwise.
The data classifier device 22 comprises one or more processors, such as computing devices executing instructions provided through programming for example, configured to operate as the passive classifier 26 and the query engine 28. The passive classifier 26 predicts classification labels Y′ for the test data 34, in part based on labels Y associated with the training data 30. The size of the training data set 30 will usually be much smaller than the size of the test data set 34. The query engine 28 uses the training data 30, test data 34 (and pool data 32 if included), the labels Y of the training data, and the predicted labels Y′ from the passive classifier 26 to identify a data entry from the test data (or pool data if included) to query for its true label Y. As described in more detail below, the query engine 28 bases the determination of which data entry to include in an augmented training data set on the predicted label values Y′ of the test data (or pool data if included) and the influence of each of the data set entries on other entries. The query engine 28 identifies the additional data entry (and its corresponding label) to be added to the training data 30 so that it may be used by the passive classifier 26 for further predictions of classification labels.
The eventual output from the data classifier device 22 is a data set schematically shown at 36 that provides a set of classification labels for at least some of the entries in the test data set 34. The type of data and classification labels may vary depending on the needs of a particular implementation.
One example scenario in which such classification could be useful is in the context of telecommunication big data analytics, such as predicting subscriber churn from a telecommunication network. Network operators are interested in the ability to predict which subscribers will churn in an effort to detect and retain potential churners before they unsubscribe from the network operator. In some cases the classifier device 22 may be employed using input training data 30 in the form of measurements for each subscriber, possibly with a labeling function indicating which subscriber has churned. The classifier device 22 acquires labels for more data points and builds a prediction (or a hypothesis) indicating if another set of subscribers (represented by the test data set 34) will churn. The labels indicating the built prediction are part of the output schematically shown at 36.
Another possible use of the classifier device 22 is to predict a likelihood that a customer Set-Top Box (STB) or Customer Premises Equipment (CPE) will fail. In many situations there will be limited data available regarding actual failure as confirmation might require lab examination of the device. The test data set 30 may include parameter measurements obtained from the STBs or CPEs and a binary labeling function of ‘failed’ or ‘OK’. The classifier device 22 is useful for learning a labeling of a very large set of STBs or CPEs by using only limited labeling resources to create a prediction model that minimizes the test set error. The resulting prediction model can be used to predict the likelihood of a customer device failure in the near future. This may allow the operator to take proactive actions to address any devices having a high likelihood to fail without having to wait for an actual failure or a customer call.
Other possible uses for the device 22 will become apparent to those skilled in the art who have the benefit of this description.
The manner in which the example device 22 builds a prediction for data classification is summarized in the flowchart diagram 40 of
The passive classifier receives data from the data storage 24 including at least training data 30 and test data 34. Some examples will also include pool data 32. At 42, the passive classifier 26 constructs a proximity graph from the data. The proximity graph includes a node for each data set entry (e.g., each measurement) and edges between nodes.
to form the final edge weights.
In this example, there is no edge shown between the nodes at 54 and 60 because neither is one of the K-closest nodes to the other. Diluting edges in this manner may reduce the processing time without penalty because there is not enough similarity between the nodes 54 and 60 to consider the edge between them for the propagation of test data set labels.
Returning attention to
In this example, the passive classifier 26 creates a characteristic vector ν such that νi=Yi if i is a training sample index, otherwise νi=0. The passive classifier applies the weights matrix W, whose entry (i,j) is formulated as described above, to the vector ν. The resulting vector {tilde over (ν)} from the product {tilde over (ν)}=Wν can be used again in the same product after reinitializing {tilde over (ν)} on the training indices to the Y values. This process of applying products of W to the characteristic vector is repeated t times in this example, where t is a user-defined parameter. The passive classifier 26 then modifies the weights of the proximity graph edges based on the propagated labels at 72. The classification algorithm uses the current propagated values to redefine the weights in the graph in the following way
where σ1 and σ2 are user-defined parameters that may or may not depend on Xi and Xj. The modified weights may then be used for additional label propagation in the step at 70. As schematically represented in
At 74 the passive classifier 26 transfers the data set with the propagated labels to the query engine 28. The newly generated labels for the test data set 34 (or the pool data set 32, if one is included) are given as a subset of the entries in the vector {tilde over (ν)} and are transferred to the query engine 28 as Y′1+1, . . . Y′N.
At 76 the query engine sorts the labels from the test data 34 (or the pool data 32, if included) to identify the label value that has a preselected characteristic. In this example, the query engine 28 identifies the minimum label value. The query engine 28 in this example sorts the Y's (i.e., the labels received from the passive classifier 26) corresponding to the pool data set 32 (if one is included) or the test data set 30. If some Y's have 0-values the query engine 28 chooses from them the corresponding data point that has the maximal sum of weights to all its neighbors: di=Σjwij. Otherwise the query engine 28 chooses the data point that has the minimal Y′ value.
At 78 the query engine 28 adds the data entry that corresponds to the label identified at 76 and the corresponding true label to the training data set 30, which can be denoted as (X1+1,Y1+1). In one example, the query engine 28 obtains the true label from a teacher (not illustrated) that provides the true label using a known technique. The query engine 28 provides the addition to the training data set or the augmented training data set to the passive classifier 26 so that the process summarized at 42 and 70-74 may be repeated by the passive classifier 26 using the updated training data. This augmentation of the training data set may be repeated a preselected number of times.
The eventual result is the set of test data 34 and the associated labels Y′ at 36 (
One way in which the example classifier device 22 differs from previous active classifiers is that it includes the process of propagating test data labels in a proximity graph to nodes that correspond to the test data (and pool data if included). Having that propagation based on the similarity between nodes connected by an edge of the proximity graph and on the label enhances the ability to more quickly converge to a solution. Increasing the training data in this way also enhances the ability of the example classifier to provide accurate results.
Another way in which the example classifier device 22 differs from previous classifiers is that the query engine identifies the data entry to add to the training data set based on the predicted label values of the pool data or test data and the influence that each node has on its neighbors in the proximity graph instead of only using a computation based on a predetermined criteria.
The combination of features described above may provide the ability to significantly reduce the time it takes to predict labels for a large data set and it may reduce the amount of computing resources required to process the information. Instead of waiting weeks or months for results, which may be the case with previous classification techniques, the example device and method described above can provide results in minutes. Additionally, such results may be obtained on much larger data sets and the results are more accurate than those possible with some previous classification techniques.
The preceding description is illustrative rather than limiting in nature. Variations and modifications to the disclosed examples may become apparent to those skilled in the art that do not necessarily depart from the essence of the contribution to the art provided by the disclosed embodiments. The scope of legal protection can only be determined by studying the following claims.
