Classifying Game Reviews by Using Natural Language Processing and Support Vector Machines with SMOTE-Tomek Algorithm

Online reviews of games are an important reference for social media users to make purchasing or downloading decisions. Such kind of electronic word-of-mouth (eWOM) provides people with important information. Thus reviews can maintain a good image and profit for game companies. Marketing managers can benefit from monitoring online reviews to understand product benefits or problems. Consumers can also use online reviews to understand the true meaning of the content. However, when facing class imbalance problems, classifiers tend to have a very high accuracy on the majority class, but an unacceptable error on the minority class which often is more important than the majority class. Therefore, this study aims to build an effective classifier to identify positive (recommendations) or negative (real thoughts) reviews by using natural language processing (NLP) and support vector machine (SVM) from real game comments. This study also used SMOTE and SMOTE-Tomek algorithms to deal with class balance problems. The results show that SVM with SMOTE-Tomek has the highest accuracy (98.52%), and the performance of SVM model is better than DT model. The results of the study can be used as a recommendation for game companies or game players.