Game Recommendation System (Steam)

Hybrid recommendation system combining content-based filtering (TF-IDF + cosine similarity) and collaborative filtering (SVD) to generate top-K personalized game recommendations on the Steam platform.

The project focuses on improving game discovery and user engagement by leveraging both item similarity and user–item interaction patterns, while addressing common recommender challenges such as cold start and sparse user behavior.

---

Highlights

• Hybrid recommender: content-based + collaborative filtering
• Cold-start handling via content similarity
• Evaluated using RMSE and ranking-based metrics
• Emphasis on product relevance and recommendation quality, not just model accuracy

---

Dataset

• Source: Steam games and user interaction data
• Entities: users, games, ratings / interactions
• Key fields:
  • user_id
  • app_id
  • title
  • genres / tags
  • description
  • price
  • rating / interaction signal

Preprocessing

• Removed records with missing critical fields (ratings, titles)
• Cleaned and normalized textual fields
• Scaled numerical attributes (e.g., price, engagement ratios)
• Prepared user–item interaction matrix for collaborative filtering

---

Approach

1. Content-Based Filtering

• Applied TF-IDF vectorization on game titles, descriptions, and tags
• Computed cosine similarity between games
• Recommended games with the highest similarity scores to a given item

This approach enables recommendations even when user interaction data is sparse or unavailable.

---

2. Collaborative Filtering

• Built a user–item interaction matrix
• Applied Singular Value Decomposition (SVD) using the Surprise library
• Learned latent factors representing user preferences and game characteristics
• Predicted unseen user–item ratings to generate personalized recommendations

---

3. Hybrid Strategy

• Combined collaborative predictions with content-based similarity
• Used content-based recommendations as a fallback for cold-start users or games
• Balanced personalization accuracy with recommendation coverage

---

Evaluation

Metrics Used

• RMSE (Root Mean Squared Error):
  Evaluates rating prediction accuracy for collaborative filtering models.
• Ranking-based evaluation (Top-K):
  • Precision@K
  • Recall@K
    These metrics better reflect real-world recommendation quality by focusing on the relevance of the top recommended items.

Model Comparison

Model	RMSE	Notes
Content-Based	—	Item similarity only
SVD (Collaborative)	✓	Strong personalization
Hybrid	✓	Best balance of relevance and coverage

---

Project Structure

Game_Recommendation_System/
├── notebooks/
│   ├── Preprocess.ipynb
│   └── Main.ipynb
├── Dataset.txt
├── README.md

---

Key Insights

• Hybrid approaches outperform standalone models by balancing accuracy and coverage
• Collaborative filtering performs well for active users but struggles with cold-start scenarios
• Content-based similarity provides robust fallback recommendations when interaction data is sparse

---

Limitations

• Offline evaluation only; no online A/B testing
• Relies on explicit or proxy interaction signals
• Label sparsity affects collaborative filtering performance

---

Next Steps

• Add implicit feedback signals (e.g., playtime, clicks)
• Introduce ANN-based similarity search (FAISS) for scalability
• Deploy a lightweight API or Streamlit demo for interactive recommendations
• Extend evaluation with diversity and novelty metrics

---

Tech Stack

• Python
• Pandas, NumPy
• Scikit-learn
• Surprise (SVD)
• TF-IDF
• Matplotlib / Seaborn