Feature selection for text classification of a news flows based on topical importance characteristic

The paper presents an approach for ranking the most valuable features for text classification task. The introduced Topical Importance Characteristic leverages the feature selection method comprising the information about the distributions of words or phrases among the topics. We compare this method to well-known TF-IDF approach and use the introduced word-ranking scheme in two classifiers: Random Forrest and Multinomial Naïve Bayes. The Accuracy of classification results was tested in the “20-Newsgroups” dataset. The developed approach outperforms TF-IDF-based methods and matches the Accuracy achieved by the more powerful state of the art approaches such as SVC on the same dataset.