G-HCF: Product Recommendation by GNN Based Hybrid Collaborative Filtering

In general, product recommendation by collaborative filtering approaches aim at learning an optimum mapping between customer and product representations. The main input they use is the customer-product matrix. These approaches, at best, utilize the customer-customer and product-product relations indirectly. The proposed method uses a customer-product graph, enriched by the customer-customer and product-product relations, with graph neural networks to predict the customer-product relations. These relations are extracted from data by matrix decomposition. Comparative assessment on real data showed improved performance compared to pure matrix decomposition and standard graph models.