A New SVM Approach to Multi-instance Multi-label Learning

In this paper, we address the problem of multi-instance multi-label learning (MIML) where each example is associated with not only multiple instances but also multiple class labels. In our novel approach, given an MIML example, each instance in the example is only associated with a single label and the label set of the example is the aggregation of all instance labels. Many real-world tasks such as scene classification, text categorization and gene sequence encoding can be properly formalized under our proposed approach. We formulate our MIML problem as a combination of two optimizations: (1) a quadratic programming (QP) that minimizes the empirical risk with L2-norm regularization, and (2) an integer programming (IP) assigning each instance to a single label. We also present an efficient method combining the stochastic gradient decent and alternating optimization approaches to solve our QP and IP optimizations. In our experiments with both an artificially generated data set and real-world applications, i.e. scene classification and text categorization, our proposed method achieves superior performance over existing state-of-the-art MIML methods such as MIMLBOOST, MIMLSVM, M³MIML and MIMLRBF.