Incorporating Prior Knowledge into Boosting for Multi-Label Classification XiaoWang

Multi-label learning deals with the problem where each instance may belong to multiple labels simultaneously. The task of the learning paradigm is to output the label set whose size is unknown a priori for each unseen instance, through analyzing the training data set with known label sets. Existing multi-label learning algorithms are almost based on the purely data-driven method. The larger the training dataset, the better the performance of the classifier. However, in some cases, training dataset is too small to obtain an accurate model, while there are some prior knowledge available. In this paper, a novel boosting based multilabel learning algorithm called KnowBoost.MH is proposed. It is derived from the famous AdaBoost.MH algorithm by incorporating prior knowledge into boosting to compensate for the lack of training data. Experimental results on two real-world multi-label datasets show that KnowBoost.MH outperforms AdaBoost.MH and some of wellestablished multi-label learning algorithms especially in the case of the lack of training data.