Feature Set Reduction for Document Classification Problems

With a growing amount of electronic documents available, there is a need to classify documents automatically. In growing text classification applications, important-term selection is a critical task for the classifier performance. Although many different techniques and heuristics have been developed, this paper shows that many of them are just a sub-set of more advanced methods originating in the field of pattern recognition. The paper puts these techniques into the pattern recognition context. It also shows that despite of some theoretical problems in this area, which are identified and described, pattern recognition techniques might be found useful for text classification tasks. The performance of different feature set reduction techniques is measured by classification accuracy when different numbers of features are selected/extracted. Results for different numbers of features and various techniques are then compared and analysed.