Multi-label Classification: a survey

Abstract

Wide use of internet generates huge data which needs proper organization leading to text categorization. Earlier it was found that a document describes one category. Soon it was realized that it can describe multiple categories simultaneously. This scenario reveals the use of multi-label classification, a supervised learning approach, which assigns a predefined set of labels to an object by looking at its characteristics. Earlier used in text categorization, but soon it became the choice of researchers for wide applications like marketing, multimedia annotation, bioinformatics. Two most common approaches for multi-label classification are transformation which takes the benefit of existing single label classifiers preceded by converting multi-label data to single label, or an adaptation which designs classifiers which handle multi-label data directly. Another popular approach is ensemble of multiple classifiers taking votes of all. Other approaches are also available namely algorithm independent and algorithm dependent approach. Based on results produced, suitable metric is used for example or label wise evaluation which depends on whether prediction is binary or ranking. Every approach offers benefits and issues like loss of label dependency in transformation, complexity in case of adaptation, improvement in results using ensemble which should be considered during design of underlying application.