Paper ‐ Diversified Complementary Product Recommendation - bryanneliu/Nature-Language-Processing GitHub Wiki

Given a query product with its product type (e.g.TV), P-Companion jointly learns the complementary product type and the particular complementary products within the targeted complementary product type subspace. Specifically, P-Companion first uses the complementary type transition module to predict diverse target complementary product types, (e.g. from TV to wall mount and cables). Then, originated from transfer metric learning, a product prediction module is developed to project the embedding of query product to each of the predicted product type subspace and successfully obtain diversified complementary products among multiple types.

Suppose the query TV product is a large, high-definition Smart TV from a popular brand. The complementary type transition module predicts wall mounts and cables as relevant complementary product types. The product prediction module then projects the TV embedding into the wall mount subspace and the cable subspace, identifying specific wall mount models and cable types that are compatible with the query TV. In this example, transfer metric learning enables the P-Companion system to leverage knowledge learned from the source domain (TV embeddings) to effectively recommend diversified complementary products within the target domain (complementary product types), enhancing the overall quality of recommendations.

Behavior-based Product Graph (BPG):

• co-purchase Bcp, co-view Bcv and purchase-afterview Bpv.
• Products as “nodes”, product types and other catalog features as “node attributes”, and pairwise item relationships as “edges”.
• Bcp − (Bpv ∪ Bcv)

P-Companion is a hierarchical multi-task learning framework, which enables the joint prediction of both complementary product types and complementary product items associated with each predicted product types.