A label compression coding approach through maximizing dependence between features and labels for multi-label classification

Label compression coding strategy aims at multi-label classification problems with high-dimensional and/or sparse label vectors. Without deteriorating classification performance significantly, its efficiency depends on two key aspects: coding raw binary label vectors into real or binary codewords shortly, and decoding binary label vectors from predicted codewords speedily, which reduce the computational costs in training and testing procedures respectively. In this paper, we propose a novel label compression coding method for multi-label classification, which maximizes dependence between features and labels using Hilbert-Schmidt independence criterion and thus considers both feature and label information simultaneously. Via solving an eigenvalue problem, our method results in a small-scale coding matrix and a fast decoding operation. The experiments on ten various bench-mark data sets illustrate that our proposed technique is superior to three existing approaches, including compressive sensing based method, principal label space transformation technique and its conditional version, according to five ranking-based and instance-based performance evaluation measures.