A novel approach for multi-label classification using probabilistic classifiers

This paper presents different approaches to solve multi-label classification. In recent study there exist many approaches to solve multi-label classification problems. Which are used in various applications such as protein function classification, music categorization, semantic scene classification, etc., It in-turn uses different evaluation metrics like hamming loss and subset loss for solving multi-label classification but which are deterministic in nature. In this work, we concentrate on probabilistic models and develop two new probabilistic approaches to solve multi-label classification. One of the two approaches is based on logistic regression and nearest neighbor classifiers. This approach is similar to binary relevance method. The training phase of this approach is same as the training phase of binary relevance method, but differs in testing phase. The second approach is based on the idea of grouping related labels. This method trains one classifier for each group and the corresponding label is called as group representative. Predict other labels based on the predicted labels of group representative. The relations between the labels are found using the concept of association rule mining.

I. Introduction

In supervised learning, most of the research concentrated on data where each example is labelled with only one label, known as single-label data. But in many real life applications, examples are associated with a subset of labels unlike single-label data. This setting, where examples are tagged with a set of labels, is known as multi-label data. Binary classification and multi-class classification are special cases of single-label classification. If the label set L contains only two elements (i. e| L | = 2) then single-label classification becomes binary classification. Otherwise (i.e.,| L |> 2), it becomes multi-class classification.