Home About Expertise Projects Blogs Contact
Recommender SystemsMatrix FactorizationCompleted

Zee Entertainment
Movie Recommender System

Building a movie recommendation engine using Pearson Correlation, Cosine Similarity, and SVD Matrix Factorization, evaluated on a 1 million+ rating dataset.

TypeRecommender Systems · Collaborative Filtering
DomainOTT / Streaming
Dataset1M+ ratings · 6,040 users · MovieLens format
ToolsPython · Surprise · Scikit-learn · Pandas
CourseScaler Academic Case Study
1M+
Ratings Processed
0.88
SVD RMSE
27%
SVD MAPE
3
CF Methods Compared
01, Business Problem

What should a Zee subscriber watch next?

Zee Entertainment runs one of India's largest OTT platforms. Recommendation quality directly drives watch time and subscriber retention. A bad recommendation wastes a user's time and increases churn. The goal: build a recommendation system that surfaces genuinely relevant content from a catalogue of thousands of titles.

🎬
The dataset
MovieLens-format data with 1M+ ratings across 6,040 users and thousands of movies. Users have rated at least 20 movies each, with an average of 166 ratings per user, a relatively dense dataset for collaborative filtering.
02, Dataset Profile

Who's in the data

AttributeValueNote
Total ratings1,000,209Dense dataset, good for CF
Unique users6,040All have rated 20+ movies
Avg ratings per user166High engagement
Max ratings per user2,314Some power users
Dominant gender~75% maleGender-biased dataset
Top age group25–34Largest rating contributor
Dominant decade1990s moviesDataset skews older films
Most common rating4 out of 5Positive bias in ratings
03, Three Recommendation Approaches

Building up from simple to complex

01
Pearson Correlation (User-User CF)
Computed Pearson correlation between user rating vectors. Recommended movies highly rated by similar users. Limited by sparsity, correlation is noisy on sparse vectors.
corrwith() · sort_values()
02
Cosine Similarity (Item-Item CF)
Built item-item similarity matrix using cosine distance. More robust than Pearson, cosine handles magnitude differences better. Produces more consistent genre-aligned recommendations.
cosine_similarity · pivot_table
03
SVD Matrix Factorization
Used Surprise library's SVD algorithm. Decomposes the user-item matrix into latent factors. Learns hidden patterns (e.g., users who like action films but haven't rated them). Best performing model.
Surprise SVD · train_test_split · accuracy.rmse
Python, svd_recommender.py
from surprise import Dataset, Reader, SVD
from surprise.model_selection import train_test_split
from surprise import accuracy

# Load data in Surprise format
reader = Reader(rating_scale=(1,5))
data = Dataset.load_from_df(df[['UserID','MovieID','Rating']], reader)

trainset, testset = train_test_split(data, test_size=0.2)

svd = SVD()
svd.fit(trainset)

predictions = svd.test(testset)
rmse = accuracy.rmse(predictions)  # RMSE ≈ 0.88

# MAPE calculation
mape = np.mean([abs((p.r_ui - p.est)/p.r_ui) for p in predictions if p.r_ui != 0])
# MAPE ≈ 0.27 → 27% average prediction error
04, Model Comparison

Which approach works best

MethodRMSEMAPEQuality
Pearson CorrelationHighHighWeak, sparse data problem
Cosine SimilarityMediumMediumBetter, genre-consistent recs
SVD Matrix Factorization0.880.27Best, latent factor learning
⭐
SVD is production-ready
RMSE of 0.88 on a 1–5 scale means predictions are typically within 1 star of the actual rating. MAPE of 27% is acceptable for sparse recommendation data. SVD should be the deployed model.
05, Tech Stack
Python 3PandasScikit-learnSurprise LibrarySVDCosine SimilarityPearson Correlation
← Back to Projects View on GitHub ↗