Problem
Production graph embeddings must balance billion-scale training, data freshness, and cold-start behavior without repeatedly rebuilding the entire representation system.
Approach
The paper combines HetSGFormer for scalable global graph learning with Incremental Locally Linear Embedding for lightweight CPU-based updates when new graph data arrives.
Public evidence
The public abstract reports an advertiser-value lift of up to 6.11% for HetSGFormer, a further 3.22% lift from the incremental module, and an 83.2% improvement in embedding refresh timeliness.
Takeaway
The two-stage design separates expensive global representation learning from targeted incremental refresh, supporting practical cold-start resilience under changing graph data.
Sources
Bibliographic details
- Authors
- Mabiao Long, Jiaxi Liu, Yufeng Li, Kefan Wang, Yi Cao, Jiandong Ding, Hao Xiong, Junchi Yan
- Venue
- IEEE Transactions on Knowledge and Data Engineering (TKDE)
- Volume / issue
- 38(11), pp. 7779–7794
- Issue date
- 2026-11 (assigned issue; online publication date not recorded here)
- Status
- Accepted; not yet recorded here as formally published
- Preprint posted
- 2025-12-15
- Accepted
- 2026-08-26