Multi-task learning for intelligent data processing in granular computing context

Abstract

Classification is a popular task in many application areas, such as decision making, rating, sentiment analysis and pattern recognition. In the recent years, due to the vast and rapid increase in the size of data, classification has been mainly undertaken in the way of supervised machine learning. In this context, a classification task involves data labelling, feature extraction, feature selection and learning of classifiers. In traditional machine learning, data is usually single-labelled by experts, i.e., each instance is only assigned one class label, since experts assume that different classes are mutually exclusive and each instance is clear-cut. However, the above assumption does not always hold in real applications. For example, in the context of emotion detection, there could be more than one emotion identified from the same person. On the other hand, feature selection has typically been done by evaluating feature subsets in terms of their relevance to all the classes. However, it is possible that a feature is only relevant to one class, but is irrelevant to all the other classes. Based on the above argumentation on data labelling and feature selection, we propose in this paper a framework of multi-task learning. In particular, we consider traditional machine learning to be single task learning, and argue the necessity to turn it into multi-task learning to allow an instance to belong to more than one class (i.e., multi-task classification) and to achieve class specific feature selection (i.e., multi-task feature selection). Moreover, we report two experimental studies in terms of fuzzy multi-task classification and rule learning based multi-task feature selection. The results show empirically that it is necessary to undertake multi-task learning for both classification and feature selection.