Multi-instance multi-label active learning

Multi-instance multi-label learning (MIML) has achieved success in various applications, especially those involving complicated learning objects. Along with the enhancing of expressive power, the cost of annotating a MIML example also increases significantly. In this paper, we propose a novel active learning approach to reduce the labeling cost of MIML. The approach actively query the most valuable information by exploiting diversity and uncertainty in both the input and output spaces. It designs a novel query strategy for MIML objects specifically and acquires more precise information from the oracle without additional cost. Based on the queried information, the MIML model is then effectively trained by simultaneously optimizing the relevance rank among instances and labels. Experiments on benchmark datasets demonstrate that the proposed approach achieves superior performance on various criteria.