Hierarchical Multilabel Classification

Computer Engineering

Multi-label classification assigns multiple labels for every instance. If labels are ordered in predefined structure called hierarchical multi-label classification (HMC). HMC is carried out using two approaches such as top down (or local) and one shot (or global). Top down approach splits the problem into sub-problems while one shot approach considers the problem as a whole. One shot approach faces a problem of explodation with huge datasets. Chain Path Evaluation (CPE) method provides solution to address this issue using either Direct Acyclic Graph or tree data structures. The method combines predictions of local classifiers based on the node’s relationships and position in the hierarchy. Probability based pruning is used by CPE and thus found to be superior when compared with other methods with respect to accuracy. This work aims at implementation of CPE and extend it further to remove label duplication. Experimental results show that removal of label duplication results in better accuracy for both MLNP and NMLNP tree based data structures.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.