Review-Aware Hybrid Feature Modeling for Enhanced Mobile App Recommendations

Abstract

With the growing volume of mobile applications, the delivery of relevant and personalized recommendations has become increasingly challenging, particularly under data sparsity and cold-start conditions. To address these issues, a review-aware hybrid recommendation framework is proposed in which user–app interactions are enriched with contextual signals derived from user-generated reviews. Sentiment- labeled feature mentions such as performance, ads and controls, are extracted from textual reviews using natural language processing and encoded into structured user and app feature matrices. These representations are integrated into a LightFM-based hybrid factorization machine model trained with a ranking-based objective to learn robust feature embeddings. The system is trained and evaluated on a filtered subset of the MobileRec dataset, with a focus on mobile game applications, a category characterized by high user engagement and diverse feedback patterns. Experimental results show that the proposed method consistently outperforms nine baselines on HIT@ $k$ and NDCG@ $k$ , with its largest relative gains under cold-start conditions. Beyond-accuracy evaluation further shows improvements in catalog coverage, intra- list diversity, novelty, and personalization. Furthermore, exploratory data analysis confirms the relevance of review semantics, revealing genre-specific user concerns and review behavior patterns. These findings demonstrate that the incorporation of feature-aware review signals improves recommendation accuracy while also enhancing interpretability and adaptability in real-world mobile app scenarios.