LSTC: When label-specific features meet third-order label correlations

Abstract

Multi-label learning (MLL) copes with applications such as text classification and image recognition where each instance is associated with multiple labels. Currently, some label-specific feature conversion approaches such as LIFT serve each individual label well, however do not take advantage of label correlation. Some label correlation exploitation approaches such as GLOCAL have independent model prediction and label correlation stages, thus limiting their learning ability. This paper introduces LSTC, a new algorithm that generates label-specific features and exploits third-order label correlation. On the one hand, for each label, we choose the same number of positive and negative representative instances inspired by density peak. These instances are used to generate a new data matrix with label-specific features. On the other hand, we train a paired output prediction network for each label based on these matrices of its own and that of two auxiliary labels. In this way, third-order label correlations are implicitly exploited. Specifically, the two auxiliary labels are the most similar and least similar, respectively, to avoid getting stuck in local optima. Experiments are conducted on seventeen benchmark datasets in comparison with ten popular algorithms. Results on eight measures demonstrate the superiority of LSTC on data from various domains except the text-domain. The source code is available at github.com/fansmale/lstc.