Learning cluster-wise label distribution for label enhancement

Abstract

Label enhancement (LE) refers to the process of recovering label distributions from logical labels for less ambiguity. Current LE techniques concentrate on learning each instance individually, which ignores the instance correlation. In this paper, we propose to learn a cluster-wise label distribution (CWLD) shared by all instances of the cluster to explore the instance correlation. The softmax-normalized sum of the CWLD and the logical label vector yields the label distribution. CWLD is learned in an iterative manner. Following instance clustering, the label distributions of all instances in each cluster are averaged. The asymmetric label correlation is then mined using heat conduction. This process is repeated until the label distribution has reached a point of convergence. Experiments were undertaken on thirteen real-world datasets compared with six state-of-the-art algorithms. Results demonstrate the effectiveness and superiority of our proposed method.