Multi-Label Learning by Instance Differentiation

Abstract

Multi-label learning deals with ambiguous examples each may belong to several concept classes simultane-ously. In this learning framework, the inherent ambigu-ity of each example is explicitly expressed in the out-put space by being associated with multiple class la-bels. While on the other hand, its ambiguity is only implicitly encoded in the input space by being repre-sented by only a single instance. Based on this recog-nition, we hypothesize that if the inherent ambiguity can be explicitly expressed in the input space appropri-ately, the problem of multi-label learning can be solved more effectively. We justify this hypothesis by propos-ing a novel multi-label learning approach named INS-DIF. The core of INSDIF is instance differentiation that transforms an example into a bag of instances each of which reflects the example’s relationship with one of the possible classes. In this way, INSDIF directly ad-dresses the inherent ambiguity of each example in the input space. A two-level classification strategy is em-ployed to learn from the transformed examples. Appli-cations to automatic web page categorization, natural scene classification and gene functional analysis show that our approach outperforms several well-established multi-label learning algorithms.