Compact learning for multi-label classification

Multi-label classification (MLC) studies the problem where each instance is associated with multiple relevant labels, which leads to exponential growth of output space. It confronts great challenge for exploration latent label relationship and intrinsic correlation between feature spaces. MLC gave rise a framework named compression (LC) obtain compact space efficient learning. Nevertheless, most existing LC methods failed consider influence or misguided by original problematic features, may result in performance degradation instead. In this paper, we present learning (CL) embed features labels simultaneously mutual guidance. The proposal versatile concept that does not rigidly adhere some specific embedding methods, independent subsequent process. Following its spirit, simple yet effective implementation called multi-label (CMLL) proposed learn low-dimensional representation both CMLL maximizes dependence embedded spaces minimizes loss recovery concurrently. Theoretically, provide general analysis different methods. Practically, conduct extensive experiments validate effectiveness method.