Two-Tower Neural Matching With Llm-Based Conversational Interface For Game Recommendation

Recommendation systems are fundamental to digital platforms, helping users discover relevant content amid information overload. This paper presents a hybrid game recommendation system that integrates a Two-Tower neural model for user-game matching with a Large Language Model (LLM) provid-ing a conversational interface. The proposed architecture leverages public Steam platform data, including game metadata and user interactions, pro-cessed through contrastive learning techniques. The Two-Tower model learns dense representations of users and games in a shared latent space, while the LLM interprets natural language preferences and translates them into struc-tured profiles for the matching model. Experimental results demonstrate the technical feasibility of the approach, achieving Precision@5 of 0.84 and NDCG@10 of 0.84, indicating effective ranking of relevant items. The sys-tem addresses cold-start challenges by enabling new users to express prefer-ences through dialogue while maintaining scalability through pre-computed item embeddings.