Reproduce “Revisiting Graph Projections for Effective Complementary Product Recommendation” on OpenResearch

About this paper

Leandro Anghinoni, Pablo Zivic, Jorge Adrian Sanchez

Researchers at MercadoLibre developed a graph projection technique for complementary product recommendations that leverages temporal directionality and higher-order relationships. This method significantly outperforms existing state-of-the-art deep learning models, achieving up to a 43% improvement in Recall@k and 38% in NDCG@k, while offering better explainability and industrial scalability.

Key insights

• Incorporating temporal directionality through inverse weighting functions (e.g., 1/k) is crucial for accurately modeling the asymmetric nature of complementary product purchases.
• Aggregating information from higher-order relationships (e.g., random walks of up to 4 steps) provides richer context for item affinity without introducing significant noise.
• A carefully designed graph projection model, tailored with domain-specific insights like temporal asymmetry, can outperform more complex deep learning architectures for complementary product recommendation while offering better explainability and scalability.

Summary by alphaXiv.