Collaborative Filtering — Movie Recommender Systems

Three assignments exploring classical and deep learning approaches to collaborative filtering on the MovieLens 100K dataset.

Projects:

1. Hybrid-cf/

Hybrid Collaborative Filtering (UCF + ICF Fusion)
A memory-based approach that fuses User-Based and Item-Based CF using cosine similarity and a weighted gamma parameter.

Hybrid NMAE: 0.1785 — outperforms both UCF (0.1860) and ICF (0.2242)

2. Jncf-deep/

J-NCF: Joint Neural Collaborative Filtering
A deep learning model with dual-tower architecture trained using CrossEntropyLoss over 5-fold cross-validation.

Average 5-Fold Test Accuracy: 46.11%

Dataset

MovieLens 100K — 100,000 ratings from 943 users on 1,682 movies, with 5 pre-defined 80/20 train/test splits.

3. 1bit-matrix-completion/

1-Bit Matrix Completion via FISTA + SVT
Convex relaxation of an NP-hard binary matrix recovery problem using logistic surrogate loss and proximal gradient methods with Nesterov acceleration.

Converged in 37 iterations | Test Sign Accuracy: 90.11% | Exact rank-5 recovery

Tech Stack