Ensemble-Based Instance Relevance Estimation in Multiple-Instance Learning

The objective of Multiple-instance learning (MIL) is to learn a mapping function from weakly labeled training data, the training data in MIL is arranged in the form of labeled bags, and every bag holds several instances. The label of the bag depends upon the characteristics of unlabeled instances. This formulation has been used in decision-making applications, such as medical image classification and molecular activity prediction. This data formulation leads to a complex hypothesis, and many existing MIL algorithms are not robust to complex hypothesis space. To deal with this limitation, this paper proposes a Fisher vector-based stacking ensemble design with an instance relevance estimation process, called relevance-based multiple-instance Fisher vector encoding (RMI-FV). The ensemble design builds on top of the instance relevance estimation mechanism. The instance relevancy calculation process employs a Gaussian mixture-based subspace clustering approach, which helps to identify instances with higher relevance to the bag label. The experiments show that the proposed RMI-FV achieves better performance than state-of-The-Art MIL approaches.