BOMGraph: Boosting Multi-Scenario E-commerce Search with a Unified Graph Neural Network

The paper tackles the sample‑sparsity issue in single‑scenario e‑commerce search by introducing BOMGraph, a unified graph neural network that jointly models text search, image‑based search, and similar‑item search.

It builds a large heterogeneous graph comprising query, trigger‑item, and product nodes across three scenarios, with three types of edges (click, co‑occurrence, similarity) and employs intra‑ and inter‑scenario meta‑paths to guide information propagation.

A heterogeneous graph encoder based on Graph Attention Networks aggregates the meta‑path‑guided sub‑graphs to generate node embeddings.

A disentangled representation module splits product embeddings into scenario‑specific and shared components, using orthogonal transformations and alignment losses to capture both commonality and differences among scenarios.

Contrastive learning with cross‑scenario data augmentation mitigates long‑tail item sparsity by pulling together similar items from other scenarios and pushing apart negative samples.

Offline experiments on three search scenarios show BOMGraph‑SKV consistently outperforms baselines (e.g., LasGNN, GraphSAGE) in NDCG, HR, and MRR; online A/B tests reveal higher CTR and RPM while maintaining comparable CVR.

The results demonstrate that unified multi‑scenario graph modeling effectively improves recall performance for e‑commerce search tasks.