Optimal Feature Selection for Learning-Based Algorithms for Sentiment Classification

Abstract

Sentiment classification is an important branch of cognitive computation—thus the further studies of properties of sentiment analysis is important. Sentiment classification on text data has been an active topic for the last two decades and learning-based methods are very popular and widely used in various applications. For learning-based methods, a lot of enhanced technical strategies have been used to improve the performance of the methods. Feature selection is one of these strategies and it has been studied by many researchers. However, an existing unsolved difficult problem is the choice of a suitable number of features for obtaining the best sentiment classification performance of the learning-based methods. Therefore, we investigate the relationship between the number of features selected and the sentiment classification performance of the learning-based methods. A new method for the selection of a suitable number of features is proposed in which the Chi Square feature selection algorithm is employed and the features are selected using a preset score threshold. It is discovered that there is a relationship between the logarithm of the number of features selected and the sentiment classification performance of the learning-based method, and it is also found that this relationship is independent of the learning-based method involved. The new findings in this research indicate that it is always possible for researchers to select the appropriate number of features for learning-based methods to obtain the best sentiment classification performance. This can guide researchers to select the proper features for optimizing the performance of learning-based algorithms. (A preliminary version of this paper received a Best Paper Award at the International Conference on Extreme Learning Machines 2018.)