Graph Neural Network Based Content Recall and Popularity Bias Mitigation for Alibaba's Home‑Decor Platform

The article describes the practical research conducted on the "Mei Ping Mei Wu" home‑decor platform, focusing on the challenges of sparse user behavior and strong popularity bias in content recommendation.

Two recall paradigms are introduced: a traditional U‑I‑I (user‑item‑item) recall that leverages user attributes, content style, space, and attached products, and a deep U‑I approach that uses vector similarity to increase diversity.

To address data sparsity, the authors propose deep side‑information mining with a Graph Neural Network (GNN). A content‑content homogeneous graph is built where each node represents a piece of content and its side information (items, categories, style, space) is used as node attributes. The pipeline includes GraphSAGE‑style sampling, an encoder that concatenates ID embeddings and aggregates neighbor vectors (mean aggregation), a decoder that reconstructs edge weights, and a binary cross‑entropy loss.

For richer representations, a multi‑view GNN framework (M2GRL) is employed. Separate intra‑view graphs (e.g., category‑category, item‑item) are learned alongside an inter‑view graph that connects content with its side information. Multi‑task learning balances the intra‑view and inter‑view objectives.

Popularity bias is mitigated using the DICE (Disentangling Interest and Conformity with Causal Embedding) framework. User and item vectors are decoupled into interest and conformity components, trained on datasets split by popularity. Four auxiliary tasks—Conformity Modeling, Interest Modeling, Click Estimation, and Discrepancy regularization—are jointly optimized.

Offline, top‑k candidate items are pre‑computed from recent 15‑day click logs, and noisy edges are filtered by a relevance threshold. Online A/B tests show consistent gains: +0.75% to +0.80% in pCTR, +0.48% to +1.50% in ipv_pCTR, and modest improvements in diversity metrics.

The authors conclude that GNN‑based high‑order relation modeling and DICE‑based bias removal effectively improve recall accuracy and diversity for low‑activity users, and outline future directions such as long‑tail node handling, GNN‑ranking integration, and multi‑factor decoupling.