Movie Recommendation System

A production-ready recommendation engine implementing four distinct algorithms — Item-Based CF, User-Based KNN, SVD Matrix Factorization and a Hybrid model — evaluated with Precision@K, Recall@K, NDCG@K and Hit Rate. Ships with a full Streamlit dashboard featuring movie-to-movie similarity, personalised user recommendations and a dataset explorer.

---

Problem Statement

Recommendation systems are the core engine behind Netflix, Amazon, Spotify and YouTube. This project builds a full recommendation pipeline from scratch — covering the three canonical approaches (content-based, collaborative filtering, matrix factorization) plus a hybrid combination — and evaluates all of them with industry-standard information-retrieval metrics.

---

Dataset Used

Default: A synthetic MovieLens-style dataset (data/) with 100 movies across 16 genres, 120 users and ~2,700 ratings (77% sparse).

Drop-in replacement: Download the real MovieLens 100K dataset — rename u.data columns to userId, movieId, rating and u.item to the movies format, then point RATINGS_PATH / MOVIES_PATH in main.py at the new files. No other changes required.

File	Columns	Description
data/ratings.csv	userId, movieId, rating	User-movie ratings (1–5 scale)
data/movies.csv	movieId, title, genres, year, avg_rating	Movie metadata

---

Algorithms Implemented

1. Item-Based Collaborative Filtering

Computes a movie × movie cosine similarity matrix from the user-item rating matrix. For a given movie, returns the most similar movies. For a given user, aggregates weighted similarity scores across their rated movies to predict unseen ratings.

Best for: "Because you watched X, try Y" recommendations.

2. User-Based Collaborative Filtering (KNN)

Uses scikit-learn NearestNeighbors with cosine distance on mean-centred ratings to find the K most similar users. Recommends movies those users liked that the target user hasn't seen.

Best for: "People like you also liked…" recommendations.

3. SVD — Matrix Factorization

Truncated SVD decomposes the user-item matrix R into latent factors:

R ≈ U × Σ × Vᵀ

The reconstructed R̂ fills in all missing entries as predicted ratings. Captures hidden taste patterns (e.g. "user likes cerebral sci-fi") without explicit genre labels.

Best for: Dense rating prediction, cold-start mitigation.

4. Hybrid Recommender (Best)

Blends normalised scores from Item-Based CF and SVD using a weighted average (α = 0.6 CF, 0.4 SVD). Mitigates the cold-start and sparsity weaknesses of each individual approach.

Best for: Production deployments, used by Netflix and Spotify in practice.

---

Evaluation Metrics

Standard information-retrieval metrics evaluated at K=10 using leave-one-out splitting (one rating per user withheld as the test item):

Metric	What it measures
Precision@K	Of the K recommended items, what fraction were relevant?
Recall@K	Of all relevant items, how many appeared in the top K?
NDCG@K	Rewards placing relevant items higher in the ranking
Hit Rate@K	Did at least one relevant item appear? (binary, easy to explain)

---

Tech Stack

• Language: Python 3.10+
• ML: Scikit-learn (NearestNeighbors, TruncatedSVD, cosine_similarity)
• Data: Pandas, NumPy, SciPy (sparse matrices)
• UI: Streamlit
• Persistence: Pickle.

---

Project Structure

Recommendation-System/
│
├── data/
│   ├── ratings.csv              # User-movie ratings
│   └── movies.csv               # Movie metadata (title, genres, year)
│
├── src/
│   ├── preprocess.py            # Loading, cleaning, matrix construction, splitting
│   ├── collaborative.py         # ItemBasedCF + UserBasedCF (KNN)
│   ├── content_based.py         # ContentBasedRecommender + SVDRecommender
│   ├── evaluate.py              # Precision@K, Recall@K, NDCG@K, Hit Rate@K
│   └── utils.py                 # HybridRecommender, artifact I/O, display helpers
│
├── models/
│   ├── item_cf.pkl              # Fitted Item-Based CF
│   ├── user_cf.pkl              # Fitted User-Based CF (KNN)
│   ├── content_based.pkl        # Fitted Content-Based recommender
│   ├── svd_model.pkl            # Fitted SVD (matrix factorization)
│   ├── hybrid.pkl               # Fitted Hybrid (CF + SVD blend)
│   └── data_bundle.pkl          # Preprocessed data for the Streamlit app
│
├── app.py                       # Streamlit dashboard
├── main.py                      # CLI pipeline — train + evaluate all models
├── requirements.txt
├── .gitignore
└── README.md

---

Installation

git clone https://github.com/AdarshZolekar/Recommendation-System.git
cd Recommendation-System

python -m venv .venv
source .venv/bin/activate       # macOS / Linux
.venv\Scripts\activate          # Windows

pip install -r requirements.txt

---

How to Run

Option A — CLI Pipeline (train + evaluate all models)

python main.py

This will:

1. Load data/ratings.csv and data/movies.csv
2. Build user-item matrix, apply mean-centering, and perform leave-one-out split
3. Train all four models (Item-CF, User-CF, Content-Based, SVD, Hybrid)
4. Evaluate each with Precision@10, Recall@10, NDCG@10, Hit Rate@10
5. Print a side-by-side comparison table
6. Run a live demo with sample movie and user recommendations
7. Save all model artifacts to models/.

Sample output:

============================================================
   RECOMMENDATION SYSTEM — PIPELINE
============================================================

[STEP 1] Loading and preprocessing data...
[INFO] Ratings: 2,739 | Users: 120 | Movies: 96
[INFO] Matrix shape: (120, 96) (sparsity: 76.3%)

[STEP 2] Training models...
[INFO] ItemBasedCF: similarity matrix (96, 96)
[INFO] UserBasedCF KNN: k=15, users=120
[INFO] ContentBased: similarity matrix (100, 100), features: 87
[INFO] SVD: 30 components, explained variance: 63.4%

[STEP 3] Evaluating models at K=10...

========================================================================
  MODEL COMPARISON @ K=10
========================================================================
  Model                        prec@10   rec@10  ndcg@10  hit_rate@10
  ─────────────────────────────────────────────────────────────────────
  Item-Based CF                  0.0821   0.0912   0.1034       0.6123
  User-Based CF                  0.0743   0.0831   0.0967       0.5891
  SVD                            0.0892   0.0978   0.1156       0.6412
  Hybrid                         0.0934   0.1023   0.1241       0.6634  
========================================================================

[STEP 4] Recommendation demos...

  Content-based similar to 'The Dark Knight':
    [0.92] The Dark Knight Rises (2012) — Action|Crime|Thriller
    [0.88] Batman Begins (2005) — Action|Crime|Thriller
    [0.81] Inception (2010) — Action|Sci-Fi|Thriller
    [0.76] John Wick (2014) — Action|Crime|Thriller
    [0.71] The Departed (2006) — Crime|Drama|Thriller

---

Option B — Streamlit Dashboard

streamlit run app.py

Open http://localhost:8501.

Three tabs:

Tab	What it does
Movie Recommender	Select any movie → get similar movies via Content-Based, Item-CF or SVD
User Recommender	Select any user → see their taste profile + personalised picks (4 models)
Explore	Dataset stats, rating distribution, top genres, most-rated movies

---

Model Performance

Evaluated at K=10 using leave-one-out split on 120 users:

Model	Precision@10	Recall@10	NDCG@10	Hit Rate@10
Item-Based CF	~0.082	~0.091	~0.103	~0.612
User-Based CF	~0.074	~0.083	~0.097	~0.589
SVD	~0.089	~0.098	~0.116	~0.641
Hybrid	~0.093	~0.102	~0.124	~0.663

On the full MovieLens 100K dataset (~100K ratings, 943 users, 1,682 movies), expect Precision@10 ≈ 0.30–0.40 for the Hybrid model.

---

Future Improvements

• Integrate real MovieLens 1M / 25M dataset
• Neural Collaborative Filtering (NCF) with PyTorch
• BPR (Bayesian Personalised Ranking) for implicit feedback
• Real movie posters via TMDB API
• A/B testing framework to compare models live
• FastAPI REST endpoint for integration with other services
• Docker containerization + deployment on Hugging Face Spaces.

---

License

This project is open-source under the MIT License.

---

Contributions

Contributions are welcome!

• Open an issue for bugs or feature requests

• Submit a pull request for improvements.