DyHAN: Dynamic Heterogeneous Graph Embedding with Hierarchical Attention

Background: Traditional graph embedding methods such as node2vec, GCN, GraphSAGE and GAT assume static graphs, which cannot capture users' evolving interests over time.

Dynamic graph representation learning (Dynamic Graph Embedding) captures both structural and temporal information; existing works focus on homogeneous dynamic graphs.

We propose DyHAN, a dynamic heterogeneous graph embedding algorithm based on hierarchical attention, which models node‑level, edge‑level and temporal‑level interactions.

Graph construction uses user behavior logs to build a heterogeneous graph with two node types (users, items) and three edge types (click, inquiry, order) across daily time slices.

The model consists of three attention layers: node‑level attention aggregates neighbor information, edge‑level attention weights different edge types, and temporal‑level attention (scaled dot‑product) aggregates across time slices; each layer can be extended with multi‑head mechanisms.

Training uses cross‑entropy loss on the final time slice; experiments on public datasets and Alibaba’s own data show DyHAN outperforms static baselines (DeepWalk, metapath2vec, GraphSAGE, GAT) and dynamic baselines (DynamicTriad, DySAT) in edge‑prediction tasks.

Online deployment in Alibaba International’s recommendation engine increased coverage and conversion rates (e.g., L‑AB +3.54%, D‑O +2.57%).

Conclusion: DyHAN advances heterogeneous dynamic graph modeling, but computational cost of per‑slice processing remains a challenge for future work.

Reference: L. Yang et al., “Dynamic heterogeneous graph embedding using hierarchical attentions,” ECIR 2020.