Jilin Hu

Research profile

Research summary

Jilin Hu is a machine learning researcher at East China Normal University whose reviewed work centers on representation learning for spatio-temporal and urban data. The Evolve catalog holds three 2026 arXiv papers, all listing Hu in a middle author position: a confounder-aware multi-view framework for urban region representation learning, the DGCPath distribution-aware generative contrastive framework for self-supervised path representation learning, and REFINE, identified by Evolve as a KDD'26 extended paper unifying contrastive and generative trajectory representation learning through closed-loop transcription. Methodologically, the portfolio combines contrastive and self-supervised learning, diffusion-generated views, variational inference, graph neural networks, and attention-based multi-view fusion, applied to urban regions, path data, and road-network trajectories. Bibliometrically, the three reviewed papers split evenly across three adjacent machine-learning fields (each exactly one third of the field distribution), with 11 distinct coauthors recorded; recurring collaborators include Sean Bin Yang, Zongyi Xu, Hua Lu, Bin Yang, Christian S. Jensen, Ying Sun, Kristian Torp, Hao Miao, Xiangmeng Wang, Tung Kieu, and Torben Bach Pedersen. Evolve's reproducibility assessment averages 70 across the three papers, with DGCPath rated high readiness and the other two medium. Review verdicts consistently credit experimental breadth — beating 10 baselines on 3 trajectory datasets and winning across 4 datasets — while flagging thin statistical testing and, in the urban-region work, heuristic deconfounding lacking causal rigor. Because no ORCID is linked, this profile is name-resolved from public bibliographic data.

Research keywords

Highlights

• Unifying contrastive and generative trajectory representation learning via closed-loop transcription

  Per Evolve's review of this KDD'26 extended paper, REFINE delivers broad wins across 4 trajectory datasets, though the supporting theory is conditional and statistical testing remains thin.

• Distribution-aware generative contrastive learning for path representations

  DGCPath combines diffusion-generated views with Jensen-Shannon distributional contrast, beating 10 baselines on 3 trajectory datasets; Evolve rates its reproducibility high while judging the novelty incremental.

• Confounder-aware multi-view learning for urban region embeddings

  This reviewed paper adds heuristic shared-component residualization to multi-view urban embeddings with strong gains on 3 cities; Evolve notes the deconfounding lacks causal rigor and statistical backing.

Timeline

1. 2026The confounder-aware multi-view urban region representation learning paper (CURE) appeared on arXiv, listing Jilin Hu as a middle author in the machine learning / urban computing field.
2. 2026The DGCPath extended version was posted on arXiv, applying diffusion-generated views and Jensen-Shannon distributional contrast to self-supervised path representation learning.
3. 2026The REFINE extended version appeared on arXiv; Evolve's review identifies it as a KDD'26 paper unifying contrastive and generative trajectory representation learning via closed-loop transcription.

Papers in the Evolve catalog

• 2026-09-14

  When Correlations Mislead: Confounder-Aware Multi-View Urban Region Representation Learning

• 2026-09-07

  DGCPath: Distribution-Aware Generative Contrastive Framework for Self-supervised Path Representation Learning -- Extended Version

• 2026-09-07

  REFINE: Trajectory Representation Learning via Closed-Loop Transcription -- Extended Version

Frequent co-authors