Transformation of discriminative single-task classification into generative multi-task classification in machine learning context

Abstract

Classification is one of the most popular tasks of machine learning, which has been involved in broad applications in practice, such as decision making, sentiment analysis and pattern recognition. It involves the assignment of a class/label to an instance and is based on the assumption that each instance can only belong to one class. This assumption does not hold, especially for indexing problems (when an item, such as a movie, can belong to more than one category) or for complex items that reflect more than one aspect, e.g. a product review outlining advantages and disadvantages may be at the same time positive and negative. To address this problem, multi-label classification has been increasingly used in recent years, by transforming the data to allow an instance to have more than one label; the nature of learning, however, is the same as traditional learning, i.e. learning to discriminate one class from other classes and the output of a classifier is still single (although the output may contain a set of labels). In this paper we propose a fundamentally different type of classification in which the membership of an instance to all classes(/labels) is judged by a multiple-input-multiple-output classifier through generative multi-task learning. An experimental study is conducted on five UCI data sets to show empirically that an instance can belong to more than one class, by using the theory of fuzzy logic and checking the extent to which an instance belongs to each single class, i.e. the fuzzy membership degree. The paper positions new research directions on multi-task classification in the context of both supervised learning and semi-supervised learning.