Learning a Distance Metric from Multi-instance Multi-label Data

Multi-instance multi-label learning (MIML) refers to the learning problems where each example is represented by a bag/collection of instances and is labeled by multiple labels. An example application of MIML is visual object recognition in which each image is represented by multiple key points (i.e., instances) and is assigned to multiple object categories. In this paper, we study the problem of learning a distance metric from multi-instance multi-label data. It is significantly more challenging than the conventional setup of distance metric learning because it is difficult to associate instances in a bag with its assigned class labels. We propose an iterative algorithm for MIML distance metric learning: it first estimates the association between instances in a bag and its assigned class labels, and learns a distance metric from the estimated association by a discriminative analysis; the learned metric will be used to update the association between instances and class labels, which is further used to improve the learning of distance metric. We evaluate the proposed algorithm by the task of automated image annotation, a well known MIML problem. Our empirical study shows an encouraging result when combining the proposed algorithm with citation-kNN, a state-of-the-art algorithm for multi-instance learning.