Game Recommendation System Using Transformer with Remastered Feature

The rapidly growing game industry makes it difficult for players to find games that match their preferences. Conventional recommendation methods are often unsatisfactory due to a lack of personalization. This study aims to design and build a web-based game recommendation system using the content-based filtering method by leveraging a fine-tuned Transformer embedding model, all-MPNet-base-v2, to deeply analyze the textual content of games. The research methodology included data collection from the Steam API (43,900 games), text preprocessing with TF-IDF for keyword extraction, and significantly, fine-tuning the all-MPNet-base-v2 model using the Knowledge Distillation method with jina-embedding-v3 as the teacher model. A novel game series identification feature using fuzzy string matching was also implemented. The resulting embedding vectors were indexed using LanceDB and deployed in a Flask web application. The research contributions are the successful domain-specific adaptation of MPNet via Knowledge Distillation and the implementation of the series identification feature. Quantitative evaluation demonstrated the fine-tuned model's superiority, achieving substantial improvements over the baseline in MRR@10 (0.5857), MAP@10 (0.5149), and Hit Rate@3 (0.90). User Acceptance Testing (UAT) with 15 respondents showed high acceptance (92.89%). Limitations include the Steam-only dataset, potential information loss from TF-IDF, and the small UAT sample size. This study confirms that fine-tuned Transformer embeddings within a content-based framework, enhanced by Knowledge Distillation, can produce effective, accurate, and well-received game recommendations, further improved by context-aware features like series identification.