Projecthack night

Built at OpenAI Codex - Fast Hacks in Tokyo

TasteGraph

TasteGraph is a personalized movie discovery system that moves beyond fixed genre tags and generic similarity scores. The system represents users, movies, genres, actors, and directors as a Neo4j knowledge graph. In parallel, an autoencoder compresses high-dimensional movie features and user preference signals into a compact 64-dimensional latent representation. Users select several favorite films, and TasteGraph builds multiple local taste neighborhoods instead of reducing their preferences to a single average point. It then recommends nearby movies and explains each recommendation through relationship paths in the graph. Our prototype demonstrates three core steps: 1. Select a favorite movie. 2. Calculate its local neighborhood. 3. Display nearby recommendations with explainable connections. The project combines graph relationships with learned representations to create recommendations that are more personal, exploratory, and understandable. The next step is to evaluate the system with a larger dataset and retrain a commercial version using only properly authorized data.

View source
// latest_project_recording.mp4
// Project brief

TasteGraph is a personalized movie discovery system that moves beyond fixed genre tags and generic similarity scores. The system represents users, movies, genres, actors, and directors as a Neo4j knowledge graph. In parallel, an autoencoder compresses high-dimensional movie features and user preference signals into a compact 64-dimensional latent representation. Users select several favorite films, and TasteGraph builds multiple local taste neighborhoods instead of reducing their preferences to a single average point. It then recommends nearby movies and explains each recommendation through relationship paths in the graph. Our prototype demonstrates three core steps: 1. Select a favorite movie. 2. Calculate its local neighborhood. 3. Display nearby recommendations with explainable connections. The project combines graph relationships with learned representations to create recommendations that are more personal, exploratory, and understandable. The next step is to evaluate the system with a larger dataset and retrain a commercial version using only properly authorized data.

// Built with
// Watch & explore