Capturing Context-Aware Route Choice Semantics for Trajectory Representation Learning

Cao, Ji; Wang, Yu; Zheng, Tongya; Song, Jie; Guo, Qinghong; Ren, Zujie; Jin, Canghong; Chen, Gang; Song, Mingli

doi:10.1109/TKDE.2026.3708608

IEEE Transactions on Knowledge and Data Engineering

Capturing Context-Aware Route Choice Semantics for Trajectory Representation Learning

Ji Cao^1,2,*, Yu Wang^1,*, Tongya Zheng³, Jie Song⁴, Qinghong Guo⁵, Zujie Ren^2,1,†, Canghong Jin³, Gang Chen¹, Mingli Song¹

¹ College of Computer Science and Technology, Zhejiang University, Hangzhou 310027, China ² Zhejiang Lab, Hangzhou 311121, China ³ Zhejiang Provincial Engineering Research Center for Real-Time SmartTech in Urban Security Governance, Hangzhou City University, Hangzhou 310015, China ⁴ School of Software Technology, Zhejiang University, Ningbo 315100, China ⁵ Polytechnic Institute of Zhejiang University, Hangzhou 310015, China
^*Equal contribution. ^†Corresponding author.

Paper arXiv Code

Overview of the CORE framework — CORE constructs a context-aware road network from multi-granular POI semantics, models route choice behavior with an MoE-based encoder, and aggregates route-choice-aware representations for downstream trajectory tasks.

Abstract

Trajectory representation learning (TRL) aims to encode raw trajectory data into low-dimensional embeddings for downstream tasks such as travel time estimation, mobility prediction, and trajectory similarity analysis. From a behavioral perspective, a trajectory reflects a sequence of route choices within an urban environment. However, most existing TRL methods ignore this underlying decision-making process and instead treat trajectories as static, passive spatiotemporal sequences, thereby limiting the semantic richness of the learned representations.

To bridge this gap, we propose CORE, a TRL framework that integrates context-aware route choice semantics into trajectory embeddings. CORE first incorporates a multi-granular Environment Perception Module, which leverages large language models (LLMs) to distill environmental semantics from point of interest (POI) distributions, thereby constructing a context-enriched road network. Building upon this backbone, CORE employs a Route Choice Encoder with a mixture-of-experts (MoE) architecture, which captures route choice patterns by jointly leveraging the context-enriched road network and navigational factors.

Finally, a Transformer encoder aggregates the route-choice-aware representations into a global trajectory embedding. Extensive experiments on 4 real-world datasets across 6 downstream tasks demonstrate that CORE consistently outperforms 15 state-of-the-art TRL methods, achieving an average improvement of 9.20% over the best-performing baseline.

Downstream Results

CORE is evaluated on four real-world datasets across six downstream tasks. Tables report CORE's scores and relative improvements over the strongest baseline; Gain values follow the metric order in each table.

Road Label Prediction

Dataset	Macro-F1 ↑	Micro-F1 ↑	Gain
Beijing	0.9303	0.9408	16.52% / 15.68%
Chengdu	0.8574	0.9009	17.97% / 8.87%
Xi'an	0.8456	0.8367	21.15% / 15.97%
Porto	0.8073	0.8551	25.98% / 19.23%

Travel Time Estimation

Dataset	MAE ↓	RMSE ↓	MAPE ↓	Gain
Beijing	3.6703	6.9956	0.2551	9.09% / 13.11% / 8.01%
Chengdu	1.3681	2.1719	0.1801	4.05% / 2.36% / 3.38%
Xi'an	2.1498	3.5783	0.1951	4.51% / 0.60% / 4.69%
Porto	1.3519	2.1151	0.2198	3.44% / 3.32% / 2.18%

Similar Trajectory Retrieval

Dataset	HR@1 ↑	HR@5 ↑	MRR ↑	Gain
Beijing	0.9611	0.9911	0.9729	5.87% / 4.97% / 5.39%
Chengdu	0.9402	0.9746	0.9554	17.64% / 7.76% / 13.82%
Xi'an	0.9540	0.9858	0.9685	7.65% / 4.84% / 6.16%
Porto	0.9746	0.9930	0.9820	14.65% / 5.66% / 7.38%

Destination Prediction

Dataset	Acc@1 ↑	Acc@5 ↑	Acc@10 ↑	Gain
Beijing	0.1076	0.2647	0.3499	16.32% / 15.39% / 13.90%
Chengdu	0.4296	0.6411	0.7189	2.90% / 2.23% / 1.86%
Xi'an	0.3631	0.5977	0.6845	2.28% / 1.49% / 1.06%
Porto	0.1734	0.3750	0.4694	3.58% / 1.46% / 1.38%

Path Ranking

Dataset	Tau ↑	Rho ↑	MAE ↓	Gain
Beijing	0.6814	0.7290	0.1283	4.77% / 3.85% / 7.43%
Chengdu	0.7619	0.8017	0.0934	0.79% / 0.78% / -0.76%
Xi'an	0.7129	0.7630	0.1007	1.49% / 1.45% / 2.61%
Porto	0.8267	0.8511	0.1103	2.87% / 2.19% / 4.67%

Trajectory Generation

Dataset	Hausdorff ↓	DTW ↓	EDR ↓	Gain
Beijing	0.4395	5.8033	0.3511	21.84% / 20.88% / 23.14%
Chengdu	0.1102	0.8371	0.2082	15.23% / 17.83% / 16.55%
Xi'an	0.2093	2.8187	0.2477	15.64% / 17.11% / 16.82%
Porto	0.4592	10.5123	0.4259	19.40% / 22.07% / 21.06%

BibTeX

@article{cao2026capturing,
  title={Capturing Context-Aware Route Choice Semantics for Trajectory Representation Learning},
  author={Cao, Ji and Wang, Yu and Zheng, Tongya and Song, Jie and Guo, Qinghong and Ren, Zujie and Jin, Canghong and Chen, Gang and Song, Mingli},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2026},
  volume={38},
  number={10},
  pages={6408-6423},
  doi={10.1109/TKDE.2026.3708608},
}