Movie Recommender System

A content-based recommendation engine built on the TMDB 5000 dataset. The system uses cosine similarity on vectorized movie metadata to surface personalized top-N recommendations in under a second.

How It Works

Raw movie data — genres, cast, crew, keywords, and plot descriptions — is cleaned and engineered into a feature-rich representation. Each movie becomes a vector in high-dimensional space, and similarity between any two films is measured via cosine distance.

Pipeline

• Data ingestion and cleaning with Pandas and NumPy — handling missing values, deduplication, and type normalization
• Feature engineering from genres, cast (top 3 billing), director, and keyword tags
• Count Vectorization to convert text metadata into numerical feature vectors
• Cosine similarity matrix computed across the full corpus
• Top-N recommendation function with sub-second query response

Key Decisions

Chose content-based filtering over collaborative filtering because the dataset lacks user rating density. The cosine similarity approach avoids the cold-start problem entirely — any movie with metadata can be recommended immediately. The modular pipeline design means swapping in a hybrid approach later requires minimal refactoring.