Learning Curves for the Analysis of Multiple Instance Classifiers

In Multiple Instance Learning (MIL) problems, objects are represented by a set of feature vectors, in contrast to the standard pattern recognition problems, where objects are represented by a single feature vector. Numerous classifiers have been proposed to solve this type of MIL classification problem. Unfortunately only two datasets are standard in this field (MUSK-1 and MUSK-2), and all classifiers are evaluated on these datasets using the standard classification error. In practice it is very informative to investigate their learning curves, i.e. the performance on train and test set for varying number of training objects. This paper offers an evaluation of several classifiers on the standard datasets MUSK-1 and MUSK-2 as a function of the training size. This suggests that for smaller datasets a Parzen density estimator may be preferrer over the other ’optimal’ classifiers given in the literature.