Constrained instance clustering in multi-instance multi-label learning

In multi-instance multi-label (MIML) learning, datasets are given in the form of bags, each of which contains multiple instances and is associated with multiple labels. This paper considers a novel instance clustering problem in MIML learning, where the bag labels are used as background knowledge to help group instances into clusters. The goal is to recover the class labels or to find the subclasses within each class. Prior work on constraint-based clustering focuses on pairwise constraints and can not fully utilize the bag-level label information. We propose to encode the bag-label knowledge into soft bag constraints that can be easily incorporated into any optimization based clustering algorithm. As a specific example, we demonstrate how the bag constraints can be incorporated into a popular spectral clustering algorithm. Empirical results on both synthetic and real-world datasets show that the proposed method achieves promising performance compared to state-of-the-art methods that use pairwise constraints.