Multiple-Instance Learning Via Random Walk

This paper presents a decoupled two stage solution to the multiple-instance learning (MIL) problem. With a constructed affinity matrix to reflect the instance relations, a modified Random Walk on a Graph process is applied to infer the positive instances in each positive bag. This process has both a closed form solution and an efficient iterative one. Combined with the Support Vector Machine (SVM) classifier, this algorithm decouples the inferring and training stages and converts MIL into a supervised learning problem. Compared with previous algorithms on several benchmark data sets, the proposed algorithm is quite competitive in both computational efficiency and classification accuracy.