PyTorch Two-Tower Neural Network Game Recommendation System: Complete Implementation from Theory to Practice

This article introduces a PyTorch-implemented two-tower neural network game recommendation system, which uses the YouTube DNN retrieval architecture and Wide & Deep ranker to build a two-stage recommendation process. The system leverages game playtime from the Steam dataset as implicit feedback, integrates multi-dimensional user and item features, effectively alleviates the cold start problem, and fully demonstrates the entire process from theory to practice.

In the era of information explosion, recommendation systems are the key bridge connecting users and content. Traditional collaborative filtering struggles with massive items and sparse data, while deep learning recommendation systems provide new ideas. The two-tower neural network architecture has become mainstream in the industry due to its efficient retrieval capability and scalability. Its core is to encode users and items into the same vector space and realize recommendations through similarity calculation. The YouTube 2016 paper first systematically elaborated this architecture, which is now widely used.

• Original Author/Maintainer: nickgreenquist
• Source Platform: GitHub
• Original Link: https://github.com/nickgreenquist/Game-Recommender-System-PyTorch-TwoTower-Model
• Release Date: May 24, 2026
• Related Projects: Sister projects to book and movie recommendation systems

This project is built based on the UCSD Steam dataset, containing about 5,437 games and 4.3 million training samples. Instead of using explicit star ratings, it uses game playtime as implicit feedback. It adopts a two-stage architecture: the two-tower model is responsible for candidate generation (filtering Top100 candidates), and the Wide & Deep ranker performs fine-grained sorting to balance recommendation quality and response speed.

User Tower Design: Builds user representation entirely based on behavior signals to alleviate cold start. It includes four history pools (like pool, dislike pool, full pool, playtime-weighted pool), as well as a genre affinity tower (based on the proportion of historical game genres) and a tag tower (TF-IDF weighted community tag vectors).

Item Tower Design: Encodes games into 128-dimensional vectors, integrating multi-dimensional features: game ID embedding (32D), genre embedding (8D), tag embedding (32D), developer embedding (12D), year embedding (8D), price embedding (4D). After feature concatenation, a two-layer projection network outputs an L2-normalized vector.

The training strategy adopts multiple industry-level best practices:

• Full Softmax Cross-Entropy Loss: Calculates the softmax cross-entropy of the entire game library at each step to provide dense gradient signals.
• Temperature Scaling: The parameter is set to 0.000977 to adjust the smoothness of softmax.
• Popularity Bias Correction: alpha=0.4 penalizes popular games and encourages exploration of long-tail content.
• Rollback Data Augmentation: Randomly shuffles user history to generate multiple training samples.
• Valve Game Filtering: Removes classic Valve games like CS:GO to avoid cross-genre recommendation contamination.

The two-stage service architecture works collaboratively:

• Retrieval Stage: The two-tower model precomputes user and item vectors, and quickly filters Top100 candidates through vector similarity search (e.g., FAISS) to ensure millisecond-level response.
• Ranking Stage: The Wide & Deep ranker uses cross features (such as interaction between user genre preferences and game genres) to perform fine-grained scoring and sorting of candidates, improving recommendation accuracy.

The project was evaluated on 2,000 unseen validation set users, using the random shuffled history protocol, where Recall@K equals the hit rate. The results show that the two-tower model can effectively recall content of interest to users, and adjusting the alpha parameter can flexibly balance popular recommendations and long-tail exploration.

Practical Insights:

1. Implicit behavior signals (such as game playtime) can replace explicit ratings to build high-quality recommendation systems.
2. ID-free embedding design alleviates cold start, allowing new users to get recommendations with a small number of behaviors.
3. Multi-dimensional feature fusion significantly improves recommendation quality.
4. The two-stage architecture balances performance and effectiveness.

Expansion Directions: Introduce sequence models to model temporal behaviors, explore self-supervised pre-training to improve representation quality, and expand multi-modal scenarios by combining visual features.