Comparative Analysis of Machine Learning Classifiers for Text Classification: Characterization of Features Probability Distribution

The proliferation of social networks, IoT, cloud computing platforms has enabled the collection and storage of massive amount of data feasible. A large portion of the gathered data is available in text form leading to requirement of efficient text mining techniques. One of the basic tasks in text mining is text classification. The paper presents a detailed comparative performance analysis of popular machine learning based classifiers for text classification. Even though support vector machine classifier outperforms others in many cases, none of the classifier can classify all datasets. In order to gain more insights, the probability distributions of features across datasets are examined and the emergence of power law probability distributions is confirmed in many. This finding provides a rationale for the effectiveness of maximum Tsallis entropy classifier recently proposed in the literature. These results will help practitioners in choosing appropriate text classification method.