A Large-Margin Approach for Multi-Label Classification Based on Correlation Between Labels

Abstract

Multi label classification is a challenging task in machine learning concerned with assigning a sample to a subset of available label set. Meaning, a sample can belong to multiple labels. Furthermore, high dimensionality of data and complex correlation between labels makes it even more interesting. For this reason, it attracted many researchers in recent years. classifier-chains (CC), one of well-known methods for multi label classification which is based on binary relevance (BR) method, incorporates label correlation by assuming an order for labels and inserting previous label outputs in feature space and achieves higher performance while still retaining relatively low time complexity. But using predicted labels as features might not be very interpretable with regards to integrating label correlation into the model, especially considering there could be different types of features in a dataset. In this paper, we propose an approach for using correlation among labels based on structure of CC by defining a large-margin model between two predicted labels. Thus directly exploiting the correlation between them in a more interpretable way. The proposed approach is evaluated using 9 multi label datasets and 2 evaluation metrics. Empirical experiments show promising results and demonstrate the effectiveness of proposed method against classifier chains algorithm.