Neo4j Movie Recommender GNN Model

Personal project more for learning than portfolio

Project Overview

Stage 1: Data Ingestion ✅

• Fetch movie data from TMDb through API.
• Map genres onto their respective IDs.
• Save full dataset.

Stage 2: Knowledge Graph Construction ✅

• Ensure Neo4j is running locally via Docker.
• Create nodes for entities (Movies, Genres), and relationships (Movies → Genres, Movies → Release Year, Movies → Popularity Scores).
• Utilize Cypher queries to ensure proper ingestion of relational data.

Stage 3: Feature Engineering ✅

• Utilize Cypher queries to extract features:
  • Shared Genres: Number of genres shared between input movies and candidate movies.
  • Shared Actors/Directors: Connections between movies via shared cast/crew.
  • Graph Metrics:
    • Node Degree: Number of direct relationships a movie has.
    • PageRank: Importance of movies in the graph structure.
• Export graph-based features from Neo4j to a pandas DataFrame.
• Label movies with a match score (0-1) for pairwise combinations.

Stage 4: GNN-Based Recommendation Model Development (Current Stage)

• Graph Construction for GNN:
  • Export nodes, edges, and adjacency matrices from Neo4j to build a graph for GNN training.
  • Use graph-based features as node attributes (e.g., popularity_bin, PageRank).
• Model Selection:
  • Build a GNN architecture using frameworks like PyTorch Geometric or DGL:
    • Graph Convolutional Network (GCN): For node embeddings.
    • Graph Attention Network (GAT): For weighing relationships between nodes.
  • Train the model for:
    • Node Classification: Classify candidate movies as recommended or not.
    • Link Prediction (Optional): Predict links between user-selected movies and candidates.
• Model Evaluation:
  • Use metrics like accuracy, F1-score (for classification), or AUC (for link prediction).

Stage 5: Recommendation System Deployment

• User Input:
  • Allow users to select 3 movies as input.
• Real-Time Feature Extraction:
  • Query Neo4j for candidate movies and their graph-based features dynamically.
• Real-Time Prediction:
  • Use the trained GNN model to recommend movies based on input.
• Deployment:
  • Deploy the pipeline with Docker and expose as an API using Flask or FastAPI.

Data Source(s)

The Movie Database: https://www.themoviedb.org/