Multi-Label Classification using Q-Learning

Abstract

Multi-label classification is an important but difficult topic that involves assigning the most appropriate subset of class labels to each document from a large label collection. The enormous label space presents a number of research obstacles, including data sparsity and scalability. In recent years, breakthrough machine learning algorithms such as tree induction using large margin partitions of the instance spaces and label vector embedding in the target space have resulted in substantial progress. Example: The input text may be a narrative document from chinastory.cn, with the labels representing storey categories that infer the possible meaning of the content. However, applying standard neural network models to the Multi-label classification problem in a haphazard manner results in sub-optimal performance because to the wide output space as well as the label sparsity problem. Despite its widespread success in other fields, Q-learning has not been investigated for multi-label classification. This paper presents the Q-learning algorithm to Multi-label classification, which was the first attempt of applying to Multi-label classification.