Multi-instance learning via instance-based and bag-based representation transformations

Recent studies show that multi-instance learning can be cast to the standard supervised learning through representation transformation in that every bag is embedded into a feature space defined by bags or instances in the training set. However, all instances from the same bag are considered to be of equal importance in the bag-based representation transformation. In this paper, we propose a new multi-instance learning algorithm by jointly considering both instance-based and bag-based representation transformations. It can be roughly divided into two steps. In the first step, the instance-based transformation is used to evaluate the importance of every instance in a bag. In the second step, the importance information is exploited to compute the weighted distances from the bag to all training bags in order to achieve the bag-based transformation. We have performed extensive experiments on several multi-instance data sets. The experimental results demonstrate the effectiveness of the proposed algorithm.