Multiple Instance Learning using Bag Distribution Parameters
David M. J. Tax · Data Archiving and Networked Services (DANS) · 2012
In pattern recognition and data analysis, objects or events are often represented by a feature vector with a fixed length. For some applications this is a severe limitation, and extensions have been proposed. One approach is Multiple-Instance Learning (MIL). Here, objects are represented by a collection of feature vectors (called a bag) and a bag is labeled positive, when at least one feature vector is member of a concept. In some situations it is not suitable to assume the presence of a concept, and the distribution of all the feature vectors in a bag is required to classify the bag. In this paper we propose a simple bag classification scheme using the parameters of the fitted distributions. Experiments show sometimes surprisingly good performances with respect to other state-of-the-art approaches.