Cost-Sensitive Reference Pair Encoding for Multi-Label Learning

A general framework for multi-label classification(MLC) called multi-label error-correcting code (ML-ECC) utilizes coding schemes in communication to improve MLC performance. The framework includes some key algorithms for some special cases of MLC, such as binary relevance and random k-labelsets. Nevertheless, current ML-ECC algorithms are usually designed for one or a few evaluation criteria, and thus may suffer from bad performance with respect to other criteria. In this paper, we propose a ML-ECC algorithm that takes the evaluation criteria into account within the error-correcting code.This algorithm, named cost-sensitive reference pair encoding(CSRPE), first transforms the MLC problem into exponentially many binary classification problems based on the criterion information and a series of reduction steps from MLC to multi-class classification and then to binary classification. The exponentially many binary classifiers cause training and prediction challenges.We resolve the training challenge by random sampling and the prediction challenge by nearest-neighbor decoding. Extensive experimental results show that CSRPE achieves stable convergence, and performs better than other ML-ECC algorithms and the state-of-the-art cost-sensitive MLC algorithms across different criteria. Furthermore, we demonstrate the potential of CSRPE in preserving the criterion information by extending it to a novel multi-label active learning algorithm. The algorithm calculates the uncertainty of each unlabeled example in the coding space of CSRPE and queries the most uncertain one. Experimental results demonstrate that the proposed algorithm is superior to existing multi-label active learning algorithms.