Netflix Recommendation System - Collaborative Filtering

A memory-based collaborative filtering recommendation system that predicts user movie ratings using both item-based and user-based approaches. The system implements cosine similarity and Pearson correlation metrics to find similar items or users, enabling personalized movie recommendations based on neighborhood-based predictions.

How to run

Clone the repository

git clone https://github.com/Khoa-BOB/Netflix_RecommendationSystem.git
cd Netflix_RecommendationSystem

Create and activate a virtual environment

python3 -m venv .venv
source .venv/bin/activate

Install dependencies

pip install -r requirements.txt

Run the program

Basic usage (with default parameters)

python main.py

Run with custom parameters

python main.py --k <neighborhood_size> --metrics <similarity_metric> --cf_type <cf_type>

Available parameters:

• --k: Neighborhood size (default: 10)
  • Example values: 3, 6, 9, 12, 15, 20
• --metrics: Similarity metric (default: cosine)
  • Options: cosine or pearson
• --cf_type: Collaborative filtering type (default: item)
  • Options: item or user

Examples:

# Item-based CF with cosine similarity and k=20
python main.py --k 20 --metrics cosine --cf_type item

# User-based CF with pearson correlation and k=10
python main.py --k 10 --metrics pearson --cf_type user

# Item-based CF with pearson and k=15
python main.py --k 15 --metrics pearson --cf_type item

Run comprehensive experiments (all combinations)

To test multiple k values across all CF types and similarity metrics:

# For local testing
python run_k_experiments.py

# For HPC cluster submission (SLURM)
sbatch submit_k_experiments.sbatch

This will test 24 configurations and generate detailed results.

---

Results Summary

ITEM-based CF with COSINE similarity

k	MSE	RMSE	MAE
3	1.537	1.240	0.933
6	1.529	1.236	0.934
9	1.534	1.238	0.937
12	1.531	1.237	0.936
15	1.530	1.237	0.935
20	1.530	1.237	0.934

Best k for Item-Cosine: k=20 (MSE: 1.530)

---

ITEM-based CF with PEARSON similarity

k	MSE	RMSE	MAE
3	1.560	1.249	0.940
6	1.552	1.246	0.934
9	1.556	1.248	0.937
12	1.555	1.247	0.938
15	1.554	1.246	0.938
20	1.552	1.246	0.937

Best k for Item-Pearson: k=20 (MSE: 1.552)

---

USER-based CF with COSINE similarity

k	MSE	RMSE	MAE
3	1.663	1.290	1.015
6	1.622	1.274	1.012
9	1.620	1.273	1.013
12	1.629	1.276	1.020
15	1.624	1.274	1.021
20	1.613	1.270	1.018

Best k for User-Cosine: k=20 (MSE: 1.613)

---

USER-based CF with PEARSON similarity

k	MSE	RMSE	MAE
3	1.635	1.279	1.002
6	1.624	1.274	0.999
9	1.626	1.275	0.999
12	1.626	1.275	0.999
15	1.626	1.275	0.999
20	1.626	1.275	0.999

Best k for User-Pearson: k=6 (MSE: 1.624)

---

Overall Best Configuration

Configuration: ITEM-based CF with COSINE similarity k: 20 MSE: 1.530 RMSE: 1.237 MAE: 0.934

---

Key Observations

1. Item-based CF outperforms User-based CF across all configurations
2. Cosine similarity performs slightly better than Pearson for both CF types
3. MSE generally improves with larger k values, with diminishing returns after k=6
4. Item-Cosine with k=20 achieved the lowest MSE of 1.530
5. User-based methods show higher error metrics (MSE ~1.6) compared to item-based (MSE ~1.5)

---

Results generated from: k_experiments_results_20251125_073304.json

Reference

Empirical Analysis of Predictive Algorithms for Collaborative Filtering