Adaptive Selection of Base Classifiers in One-Against-All Learning for Large Multi-labeled Collections

摘要

In this paper we present the problem found when studying an automated text categorization system for a collection of High Energy Physics (HEP) papers, which shows a very large number of possible classes (over 1,000) with highly imbalanced distribution. The collection is introduced to the scientific community and its imbalance is studied applying a new indicator: the inner imbalance degree. The one-against-all approach is used to perform multi-label assignment using Support Vector Machines. Over-weighting of positive samples and S-Cut thresholding is compared to an approach to automatically select a classifier for each class from a set of candidates. We also found that it is possible to reduce computational cost of the classification task by discarding classes for which classifiers cannot be trained successfully.