Multi-instance Learning Algorithm

In the 90s of last century, Dietterich et.al. firstly proposed Multi-Instance Learning concept which was originated from their research of drug activity prediction. Multiple-Instance Learning is regarded as a learning paradigm extensively exists in real world with high noisy and ambiguity. And besides supervised learning, unsupervised learning and enforcement learning, it is also categorized as the fourth learning framework because of its uniqueness.In this paper, a brief introduction of Multi-Instance Learning is given including its origin, related theoretical researches and algorithms designed. Following an analysis of some popular algorithms, a new Multi-Instance Learning algorithm named NRBF-MI is proposed based on normalized radial basis function network. This algorithm presented a new method for training the network structure; furthermore new kernels function which can deal with the labeled bags instead of points is given along with the analysis of the relation between the kernel radius and network performance. At present, most researches of Multi-Instance Learning regarded it be a learning framework closer to supervised learning, and therefore most of the algorithms of Multiple-Instance Learning were constructed through applying supervised learning method. In this paper, as an effort of introducing unsupervised learning into Multi-Instance Learning, a new algorithm for finding the multi-instance predictive structure based on the Agglomerative clustering is proposed and named as AGG-MI.Experiments on benchmark datasets MUSK and LJ-r.f.s showed that the NRBF-MI is a high efficiency algorithm for Multi-Instance Learning which has excellent prediction accuracy with simpler structure. And the analysis of the relation between the NRBF-MI's kernel radius and its performance was also approved. Experiments on these datasets also showed that AGG-MI could efficiently find the multi-instance predictive structure.