Movie Recommender System

The Problem

Recommendation systems drive an estimated 35% of e-commerce revenue and 80% of Netflix viewing hours. But most tutorials stop at "cosine similarity on a toy dataset" without exploring what actually makes recommendations good — or building something a real user can interact with.

I wanted to go deeper: build a dual-approach system that combines content understanding with collaborative signals, and ship it as a live product people can use.

The Approach

• Content-Based Filtering: TF-IDF vectorization on movie descriptions, genres, and cast, combined with cosine similarity to find movies that share characteristics with the user's favorites
• Collaborative Filtering: User-item interaction matrix factorization to surface patterns like "people who liked X also liked Y"
• Hybrid Ranking: Weighted combination of both approaches for more robust recommendations that handle cold-start gracefully
• Interactive Frontend: Streamlit interface where users select a movie and instantly see personalized recommendations with posters, ratings, and genres

Key Results

• Content-based approach captures genre/cast similarities that collaborative filtering misses
• Collaborative approach surfaces non-obvious connections between movies
• Hybrid combination outperforms either approach alone on relevance metrics
• Fully productionized with interactive web interface

Tech Stack