Building a Hybrid Movie Recommender with Neo4j and Graph Data Science

This capstone was where graph modeling and recommendation strategy clicked for me. Instead of treating recommendations as just a matrix problem, we modeled users, movies, genres, and directors as connected entities and used that structure directly.

Team Context

This was a collaborative project. Team members included:

• Peter Mangoro
• Bekithemba Nkomo
• Masheia Dzimba

Assignment Focus

We built a hybrid recommender by combining:

• collaborative edges (User -> Movie ratings)
• content structure (Movie -> Genre, Movie -> Director)
• GDS algorithms for similarity, embeddings, and community analysis

What We Built

• Graph schema with User, Movie, Genre, and Director
• Cypher + Python workflows for loading and profiling interactions
• Multiple recommendation lenses:
  • overlap similarity (Jaccard-like patterns)
  • embedding-based retrieval (FastRP + kNN)
  • community-aware perspective (Louvain)

Key Findings

• Content nodes improved explainability: recommendations could be justified by shared genres/directors, not just co-ratings.
• Data sparsity mattered: in a small dataset, pure collaborative signals were unstable for some users.
• Embedding + neighborhood approaches offered stronger behavior in sparse pockets than overlap alone.
• Community detection added a useful segmentation lens for user taste patterns.

Lessons Learned

The most useful lesson was that hybrid recommenders are not just “add more algorithms.” The graph design itself determines whether recommendations remain interpretable and robust when user behavior is sparse.

I also learned to frame model choices by business behavior (cold-start, explainability, stability), not by algorithm popularity.

Skills I Gained

• Hybrid recommender design with graph-native features
• GDS workflow design (projection, similarity, embedding, community detection)
• Evaluation of tradeoffs across sparse user-item graphs
• Collaborative project execution with clear technical handoff points

Artifacts

• Main notebook: P_Mangoro_recommender_project.ipynb