Implicit Feedback Recommendation System (Collaborative Filtering & Graph Models)

This project builds a recommendation engine for large-scale implicit user–item interaction data (Dataset 3). It generates top-20 personalized item recommendations and evaluates multiple algorithms using NDCG@20.

Overview

• Dataset size: ~52k users, ~91k items, ~2.3M interactions
• Data type: Implicit positive feedback
• Goal: Recommend 20 items per user for leaderboard submission
• Pipeline: Load → Preprocess → Train → Evaluate → Generate Submission

Implemented Models (with rationale)

Item-Based Collaborative Filtering (Cosine)

Strong, reliable baseline for sparse implicit datasets.

Item-Based CF (Jaccard)

Co-occurrence based similarity for binary interactions.

SVD Matrix Factorization

Learns latent factors for users and items.

Neural Collaborative Filtering

Models nonlinear user–item interaction patterns.

Alternating Least Squares (ALS)

Scalable matrix factorization for implicit feedback.

LightGCN

Graph-based model leveraging user–item bipartite structure.

Ensemble

Tests whether combining models yields improvements.

Best performing approach: Item-Based CF (Cosine)

Results (NDCG@20)

Model	NDCG@20
Item-CF (Cosine)	0.3056
Neural CF	0.2876
Jaccard CF	0.1635
LightGCN	0.1588
SVD	0.1803
ALS	0.0293
Ensemble	0.0050

How to Run

Evaluate a model

python recommender_system.py --mode evaluate --algorithm itemcf_cosine