Joint Classification with Heterogeneous Labels Using Random Walk with Dynamic Label Propagation

Yongxin Liao, Shenxi Yuan, Jian Chen, Qingyao Wu, Bin Li

Abstract: This paper studies a new machine learning strategy called joint classification with heterogeneous labels (JCHL). Unlike traditional supervised learning problems, JCHL uses a single feature space to jointly classify multiple classification tasks with heterogeneous labels. For instance, biologists usually have to label the gene expression images with developmental stages and simultaneously annotate their anatomical terms. We would like to classify the developmental stages and at the same time classify anatomical terms by learning from the gene expression data. Recently, researchers have considered using Preferential random walk (PRW) to build different relations to link heterogeneous labels, thus the heterogeneous label information can be propagated by the instances. On the other hand, it has been shown that learning performance can be significantly enhanced if the dynamic propagation is exploited in PRW. In this paper, we propose a novel algorithm, called random walk with dynamic label propagation (RWDLP), for the JCHL problems. In RWDLP, a joint transition probability graph is constructed to encode the relationships among instances and heterogeneous labels, and we utilize dynamic label propagation in the graph to generate the possible labels for the joint classification tasks with heterogeneous labels. Experimental results have demonstrated the effectiveness of the proposed method.

External IDs:dblp:conf/pakdd/LiaoYCWL16