MIRSVM: Multi-instance support vector machine with bag representatives

Multiple-instance learning (MIL) is a variation of supervised learning, where samples are represented by labeled bags, each containing sets of instances. The individual labels of the instances within a bag are unknown, and labels are assigned based on a multi-instance assumption. One of the major complexities associated with this type of learning is the ambiguous relationship between a bag's label and the instances it contains. This paper proposes a novel support vector machine (SVM) multiple-instance formulation and presents an algorithm with a bag-representative selector that trains the SVM based on bag-level information, named MIRSVM. The contribution is able to identify instances that highly impact classification, i.e. bag-representatives, for both positive and negative bags, while finding the optimal class separation hyperplane. Unlike other multi-instance SVM methods, this approach eliminates possible class imbalance issues by allowing both positive and negative bags to have at most one representative, which constitute as the most contributing instances to the model. The experimental study evaluates and compares the performance of this proposal against 11 state-of-the-art multi-instance methods over 15 datasets, and the results are validated through non-parametric statistical analysis. The results indicate that bag-based learners outperform the instance-based and wrapper methods, as well as MIRSVM's overall superior performance against other multi-instance SVM models, having an average accuracy of 82.6%, which is 2.5% better than the best performing state-of-the-art MI classifier.