Multilabel classification by a hierarchy
View Patent ↗A technique of extracting hierarchies for multilabel classification. The technique can process a plurality of labels related to a plurality of documents, using a clustering process, to cluster the labels into plurality of clusterings representing a plurality of classes. The technique classifies the documents and predicts a plurality of performance characteristics, respectively, for the plurality of clusterings. The technique selects at least one of the clusterings using information from the performance characteristics and adds the selected clustering into a resulting hierarchy.
1. A method for processing information, the method comprising:
providing a plurality of documents under control of a processing device, each of the plurality of documents having a label;
processing the labels related to the plurality of documents, using a clustering process, to cluster the labels into a plurality of clusterings representing a plurality of classes;
classifying the documents using the clusterings representing the plurality of classes;
predicting a plurality of performance characteristics, respectively, for the plurality of clusterings;
wherein the predicting the plurality of performance characteristics provides a performance measure to predict how accurate the clustering for classification at a given layer of a hierarchy;
selecting at least one of the clusterings using information from the performance characteristics; and
adding the selected clustering into the hierarchy.
2. The method of claim 1 wherein the labels are a current set of labels, the current set of labels are all of the labels associated with the plurality of documents;
wherein the hierarchy includes a current node;
wherein the clustering process comprises a k-means algorithm to cluster labels, the k-means algorithm being provided for each integer value of clusters; and
further comprising:
inputting through a user interface an integer value representing a number of clusters into the clustering process under control of the processing device, the integer value ranging from two (2) to a maximum number, whereupon the maximum number is an integer greater than two (2).
3. The method of claim 2 wherein the current node is the root node for a first step and not the root node for other steps; wherein the plurality of documents are stored in electronic form in at least one mass data storage device, the mass data storage device being coupled to a computing device, the computing device having a computer readable memory.
4. The method of claim 1 wherein each of the clusterings is defined as a single class; wherein the classifying comprises providing each of the documents in one or more classifications.
5. The method of claim 1 further comprising determining a hierarchy performance characteristic of the hierarchy.
6. The method of claim 1 further comprising selecting one of the clusterings having a larger performance characteristic.
7. The method of claim 1 wherein for each of the clusterings constructing a metalabel; and adding a child node to the hierarchy, the child node being associated with the metalabel.
8. The method of claim 7 wherein for each of the clusterings with a single label further comprising no meta-label is constructed.
9. The method of claim 1 further comprising iteratively repeating the processing, classifying, predicting, selecting, and adding until a condition is achieved; wherein the iteratively repeating is recursive from a preceding step; wherein the condition is achieved when each clustering consists of a single label.
10. The method of claim 1 further comprising outputting an F1 measure for the hierarchy.
11. A system for processing information, the system comprising:
a computing device;
a data storage device comprising a plurality of documents, each of the plurality of documents comprising a label, the plurality of documents being stored in electronic form, the data storage device being coupled to the computing device;
a first computer code directed to processing the labels related to the plurality of documents, using a clustering process, to cluster the labels into plurality of clusterings representing a plurality of classes;
a second computer code directed to classifying the documents;
a third computer code directed to predicting a plurality of performance characteristics, respectively, for the plurality of clusterings;
wherein the predicting the plurality of performance characteristics provides a performance measure to predict how accurate the clustering for classification at a given layer of a hierarchy;
a fourth computer code directed to selecting at least one of the clusterings using information from the performance characteristics;
a fifth computer code directed to adding the selected clustering into the hierarchy.
12. The system of claim 11 wherein the labels are the current set of labels, the current set of labels are all of the labels associated with the plurality of documents;
wherein the hierarchy includes a current node; wherein the clustering process comprises a k-means algorithm to cluster labels, the k-means algorithm being provided for each integer value of clusters; and
further comprising a user interface configured for inputting through the user interface an integer value representing a number of clusters into a clustering process under control of the processing device, the integer value ranging from two (2) to a maximum number, whereupon the maximum number is an integer greater than a minimum number.
13. The system of claim 12 wherein the current node is the root node for a first step and not the root node for other steps.
14. The system of claim 11 wherein each of the clusterings is defined as a single class; and further comprising a code directed to selecting one of the clusterings having a highest performance characteristic.
15. A computer application, the computer application being stored on a computer readable memory device, the computer readable memory device comprising:
a first computer code directed to processing a plurality of labels related to a plurality of documents, using a clustering process, to cluster the labels into a plurality of clusterings representing a plurality of classes;
a second computer code directed to classifying the documents;
a third computer code directed to predicting a plurality of performance characteristics, respectively, for the plurality of clusterings;
wherein the predicting the plurality of performance characteristics provides a performance measure to predict how accurate the clustering for classification at a given layer of a resulting hierarchy;
a fourth computer code directed to selecting at least one of the clusterings using information from the performance characteristics;
a fifth computer code directed to adding the selected clustering into the resulting hierarchy.