Graph-Based Sparse Behavior Recall Models for Content Recommendation

Background: In video recommendation, many users exhibit sparse behavior, making traditional sequence modeling insufficient. Various solutions exist, including leveraging side information, pre‑training on dense users, cross‑domain behavior transfer, and knowledge‑graph construction.

This article explores graph‑based methods to address sparse behavior by constructing user‑item graphs enriched with content knowledge graphs. It distinguishes two main categories of graph models: embedding‑based (e.g., KGE, CKE, DKN, SHINE) and path‑based (e.g., RippleNet, PER, Meta‑Graph, KGAT).

Embedding‑based models pre‑train graph embeddings separately from the recommendation task, while path‑based models feed the raw graph structure into the model for end‑to‑end learning.

We propose a novel sparse‑behavior sequence expansion technique that combines video‑to‑video (v2v) expansion with knowledge‑graph augmentation, achieving a 6% pCTR lift for sparse users.

Our Graph Attention Deep Recall Model (GADM) builds a base topology and applies multiple aggregation strategies, including RippleNet‑style attention, GCN, GraphSAGE, and Bi‑Interaction aggregators. The model incorporates KG‑guided gating and a Self‑Adaption Generative Graph Attention (SGGA) mechanism, later extended to a gated version (SGGGA) that uses encoder‑decoder attention to weight graph edges.

Experimental results show GADM‑v1 improves pCTR by 1.5% and uCTR by 2.65%; GADM‑v2 improves pCTR by 2.1% and uCTR by 1.4%.

Conclusion: Graph and knowledge‑graph methods can effectively mitigate sparse‑behavior challenges in recommendation systems, and ongoing work continues to refine these techniques.

Team: The work is from Alibaba’s Taobao "Guangguang" algorithm team, which welcomes collaborators and is recruiting talent interested in machine learning and content understanding.