The Research of Methods for Texts Classification for Unbalanced Data

This research was dedicated to analyzing methods for texts classification for unbalanced data, also the most effective of them was defined and implemented taking into account the pecularities of the Russian language. It was found that the most effective way to work with unbalanced data is to change the output threshold or classes weights rather than changing the distribution of the training data set randomly or synthetically. Thus, it was defined that the most effective approach is to use cost-sensitive methods. For empirical verification of theoretical conclusions was selected the task of classifying statements in Russian depending on whether they are correct or contain any offensive content. As a result, it was found that the usage of cost-sensitive methods did not help to improve the results of the texts classification compared to training model on moderately unbalanced data for this task. However, after conducting such experiment using strongly skewed data, it was found that the working with weighted models indeed leads to an improved classification result.