PREDICTING REVIEW HELPFULNESS IN VIDEO GAMES: A COMPARATIVE ANALYSIS OF MACHINE LEARNING MODELS AND NLP INTEGRATION

This paper investigates the prediction of video game review helpfulness on the Steam platform using machine learning and natural language processing (NLP) techniques. We applied three models—XGBoost, Extreme Learning Machine (ELM), and Ridge regression—to predict helpfulness scores as both a regression and binary classification problem. XGBoost demonstrated the best performance, while ELM offered a computationally efficient alternative. Text features generated from DistilBERT were incorporated, but their inclusion did not significantly enhance model accuracy. Our findings suggest that non-textual features, such as review length, playtime, and helpful votes, are more influential in determining helpfulness. Early predictions of review helpfulness could benefit users by highlighting valuable feedback and aiding developers in refining their games. Future research will explore fine-tuning NLP models on larger datasets and incorporating additional features, such as sentiment analysis, to improve performance.