Actor Recommendation System using Collaborative Filtering (SVD)

This project implements an Actor Recommendation System that predicts which actors a user is most likely to prefer based on their historical movie ratings.
The system uses Collaborative Filtering, specifically Singular Value Decomposition (SVD) from the Surprise library, to learn user–actor interaction patterns and generate personalized Top-N recommendations.

---

Features

• Builds a user–actor rating matrix by merging MovieLens ratings with actor metadata extracted from TMDb.
• Uses SVD matrix factorization to model latent preferences.
• Supports Top-N actor recommendations for any user.
• Evaluated using:
  • Precision@k
  • Recall@k
  • NDCG@k
• Clean workflow from data extraction → preprocessing → model training → evaluation.

---

Dataset Used

1. MovieLens Ratings Dataset

• Contains: userId, movieId, rating, timestamp
• User’s rating history forms the base interactions.

2. TMDb Actor Metadata

Collected via API:

• For each movie → list of top 5 actors
• Stored as movie_actors.csv

3. Movie–Actor Relationship Dataset

Created via preprocessing:

• movie_id_db, movie_title, actor_name
• Each row = one actor per movie
• Used to map user ratings to actors.

4. User–Actor Ratings

After merging:

• userId
• actor_name
• average_rating
  (Mean rating a user gives to movies an actor appears in.)

---

Workflow / Pipeline

MovieLens Ratings + TMDb Actors │ ▼ Create movie_actor_relationships.csv │ ▼ Merge → Build user_actor_ratings.csv (user → actor avg rating) │ ▼ Train SVD Model using Surprise Library │ ▼ Predict Ratings → Generate Top-N Actor Recommendations │ ▼ Evaluate using Precision@k, Recall@k, NDCG@k

---

Model Used — SVD (Singular Value Decomposition)

SVD factorizes the interaction matrix into:

• User latent factors
• Actor latent factors
• Bias terms (user bias + actor bias)

The model is trained via Stochastic Gradient Descent (SGD) with regularization to prevent overfitting.

---

Evaluation Metrics

Precision@10

“How many of the top 10 recommended actors were relevant?”

Recall@10

“How many relevant actors were captured in the top 10 recommendations?”

NDCG@10

“Were relevant actors ranked at the top of the recommendation list?”

Sample Output (User 5):

Precision@10 = 0.90 Recall@10 = 0.32 NDCG@10 = 0.93

Conclusion

• Very accurate recommendations (high precision)
• Limited coverage (moderate recall)
• Excellent ranking quality (high NDCG)

---

---

Installation

pip install pandas numpy scikit-learn surprise

License
This project is for educational and academic purposes.

Author
Shailja Patil
Master’s Student — CSE, IIT Kharagpur
Special focus: ML, Data Analytics, and Recommender Systems