1Solving the Multi-Instance Problem with Neural Networks

Multi-instance learning was coined by Dietterich et al. recently in their investigation on drug activity prediction. In such a learning framework, the training examples are bags composed of instances, and the task is to predict the labels of unseen bags through analyzing the training bags with known labels. A bag is positive if it contains at least one positive instance, while it is negative if it contains no positive instance. However, the labels of the instances constituting the training bags are unknown. In this paper, the open problem of designing multi-instance modification for neural networks is addressed. In detail, a neural network algorithm named Backpropagation-MIP is presented, which is derived from the popular Backpropagation algorithm through employing a new error function capturing the nature of multi-instance learning, i.e. the labels of the training bags instead of that of the training instances are known. Experiments on both the real-world and artificial benchmark multi-instance data show that the performance of Backpropagation-MIP is comparable to that of some well-established multi-instance learning methods. 1