Computer-implemented method for performing hierarchical classification
Inventors
Beskales, George • Kraemer, John • Ilyas, Ihab F. • Cleary, Liam • Roome, Paul
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11232143-B1
Publication Date
2022-01-25
Expiration Date
Abstract
Given a number of records and a number of target classes to which these records belong to, a (weakly) supervised machine learning classification method leverages known possibly dirty classification rules, efficiently and accurately learns a classification model from training data, and applies the learned model to the data records to predict their classes.
Core Innovation
The invention describes a computer-implemented hierarchical classification system that performs hierarchical class prediction using weak supervision from possibly noisy classification rules. A classification model is built from training data records having known classification labels and unlabeled records, where the model includes a plurality of per-class binary classifiers. Each binary classifier is configured to predict whether an unlabeled record belongs to a certain hierarchical class.
For each unlabeled record, hierarchical classifications are predicted using the plurality of binary classifiers, and classification probabilities are combined via chain-rule marginalization of conditional probability Pr(R∈C|R∈parent(C)). A scoring function scores the predicted hierarchical classifications, and a top-scoring predicted hierarchical classification is obtained using a search algorithm. Search termination is supported by score upper bounds based on membership probability and class specificity, and the top-scoring hierarchical class is selected for each record.
The system narrows the prediction search space by generating candidate classes for each record and by applying record deduplication or clustering to prune the search space. Candidate-class generation can be based on DNF-based blocking, locality sensitive hashing, and greedy agglomerative hierarchical clustering, where record-specific candidates restrict which hierarchical classes are considered. The system further incorporates targeted active learning and integrates rule-based labels using down-sampling or weighting and conflict resolution via human review.
Claims Coverage
The provided partial content includes one independent claim that covers a computer-implemented hierarchical classification method using multiple binary classifiers, scoring, and selection of a top-scoring hierarchical classification via a search algorithm. Dependent refinements add consolidation of unlabeled records and restriction to candidate hierarchical classes.
Hierarchical classification using plurality of binary classifiers and search-based top-scoring selection
Training on training data records with known labels and unlabeled records; building a classification model that includes a plurality of binary classifiers, each binary classifier predicting whether an unlabeled record belongs to a certain hierarchical class; predicting hierarchical classifications for each unlabeled record using the plurality of binary classifiers; scoring the predicted hierarchical classifications; obtaining a top-scoring predicted hierarchical classification for each unlabeled record using a search algorithm; and outputting the top-scoring predicted hierarchical classification for each unlabeled record.
Consolidating unlabeled records by deduplication or clustering before prediction
Consolidating unlabeled records by deduplication or clustering so that predicting, scoring, obtaining a top-scoring predicted hierarchical classification, and outputting the top-scoring predicted hierarchical classification are carried out on the consolidated records.
Restricting prediction to candidate hierarchical classes per record
Identifying candidate hierarchical classes for each unlabeled record and predicting that record’s hierarchical classification using only the identified candidate hierarchical classes.
Across the independent claim and its dependents, the core claim coverage centers on hierarchical classification formed from a plurality of binary classifiers, with predicted hierarchical classifications scored and a top-scoring hierarchical classification selected via a search algorithm. Additional coverage includes consolidation of unlabeled records by deduplication or clustering and limiting prediction to candidate hierarchical classes for each record.
Stated Advantages
Prunes the search space for hierarchical class prediction by using candidate class generation and record deduplication/clustering.
Provides targeted active learning selection of high-impact records for human labeling using entropy-based measures.
Resolves conflicts between rule-based labels and other labels via human review.
Documented Applications
No documented applications found
Interested in licensing this patent?