An Evidence-Theoretic k-Nearest Neighbor Rule for Multi-label Classification

In multi-label learning, each instance in the training set is associated with a set of labels, and the task is to output a label set for each unseen instance. This paper describes a new method for multi-label classification based on the Dempster-Shafer theory of belief functions to classify an unseen instance on the basis of its k nearest neighbors. The proposed method generalizes an existing single-label evidence-theoretic learning method to the multi-label case. In multi-label case, the frame of discernment is not the set of all possible classes, but it is the powerset of this set. That requires an extension of evidence theory to manipulate multi-labelled data. Using evidence theory makes us able to handle ambiguity and imperfect knowledge regarding the label sets of training patterns. Experiments on benchmark datasets show the efficiency of the proposed approach as compared to other existing methods.