Feature Engineering and Classifier Evaluation Using Comparative NLP for Mobile Game and Educational Application Recommendation

Abstract

Mobile application (APP) reviews published on platforms such as the Google Play Store carry rich signals relevant to recommendation prediction, yet prior work has not systematically compared individual text representation and classifier combinations in this context. The present study addresses this gap by constructing an NLP-based evaluation framework that benchmarks 15 text representations drawn from five categories: (1) bag-of-words, (2) TF-IDF, (3) word embeddings, (4) transformer-based encodings, and (5) N-gram representations— paired with five machine learning classifiers across two independent datasets: 9,664 mobile game reviews and 14,344 educational APP reviews. Class imbalance is corrected using SMOTE applied exclusively to training partitions, and model interpretability is examined through random forest (RF) feature importance scores. On the game review dataset, sentimentaugmented hybrid representations outperform purely lexical approaches, whereas BERT yields the strongest results on the educational dataset. RF achieves the highest Macro-F1 in both domains. These findings confirm that no single configuration dominates universally and underline the practical necessity of domain-sensitive evaluation strategies.