Improved Conditional Dependency Networks for Multi-label Classification

Abstract

Multi-label classification (MLC) is the supervised learning problem where an instance is associated with multiple labels, rather than with a single label. The widely known binary relevance method (BR) for multi-label classification considers each label as an independent binary problem and has been sidelined in the literature due to perceived inadequacy of label correlations. In this paper, we outline several BR-based classification methods and present our improved conditional dependency networks for multi-label classification (ICDN). ICDN inherits the framework of double layer based classifier chain (DCC) to exploit the label correlations in training stage and modifiers the conditional dependency networks (CDN) by initializing the input values of the second layer with the prediction values from the first layer during the testing stage. The main contribution of the algorithm is that it reduces randomization of input for the conditional dependency networks and improves convergence rate. Experiments on benchmark datasets demonstrate that ICDN obtains the best predictive performance across several datasets under several evaluation methods specifically designed for multi-label classification.