Multi-dimensional classification via class space fusion and comprehensive label correlations

Multi-dimensional Classification (MDC) is a new multi-output learning paradigm, where each instance in this framework is annotated with labels from multiple semantic class spaces. The vital challenge in MDC problem is how to address the heterogeneity of output space. Existing methods typically focus on the decomposition of output space, rarely consider the comprehensive fine-grained label correlations and the exploration of the intrinsic structure of the output space. And the local label correlations, which only exist within specific groups of instances, remain largely untapped in existing MDC methods. Therefore, we focus on tackling these above problems in this paper by simultaneously conducting label space fusion and considering the comprehensive label correlations to fully exploit the latent discriminative information hidden in the output space. Heuristically, we utilize low-rank representation method to achieve label space fusion to capture the intrinsic structure and implicit global correlation of label space. Then the predictive MDC model is extrapolated from the reconstructed instances equipped with the fused label space. Moreover, the topological structure information of samples is fully taken advantage of for MDC model establishment. Benefiting from these strategies, the heterogeneous label space can be fully excavated to boost the performance of MDC. Empirically, comparative experiments are conducted over several multi-dimensional benchmark data sets, and corresponding experimental results firmly illustrate that the proposed algorithm can effectively solve the MDC problem and outperform state-of-the-art methods.