Comparison of distances for multi-label classification with PCTs

Multi-label classification has received significant attention in the research community over the past few years: this has resulted in the development of a variety of multi-label classification methods. These methods either transform the multi-label dataset to several simpler datasets or adapt the learning algorithm so it can handle the multiple labels. In this paper, we consider the latter approach. Namely, we use predictive clustering trees to perform multi-label classification. Furthermore, we perform an experimental comparison of four distance measures used to select the splits in the nodes of the trees. The experimental evaluation was conducted on 6 benchmark datasets using 6 different evaluation measures. The results show that, averaged overall, the Euclidean distance and the Hamming loss yield the best predictive performance.