AutoPRL: Automated Path Representation Learning via a Learnable View Generator

Sean Bin Yang; Ying Sun; Yunyao CHeng; Hao Miao; Jilin Hu; Bin Yang; Torben Bach Pedersen; Christian S. Jensen

cs.LGOct 26, 2026

AutoPRL: Automated Path Representation Learning via a Learnable View Generator

Authors: Sean Bin Yang, Ying Sun, Yunyao CHeng, Hao Miao, Jilin Hu, Bin Yang, Torben Bach Pedersen, Christian S. Jensen

Organizations: Aalborg University, Denmark. · Chongqing University of Posts and Telecommunications, China. · East China Normal University, Shanghai, China.

Abstract

Self-supervised learning has emerged as a powerful paradigm for Path Representation Learning (PRL) in road networks. PRL aims to learn informative representations of paths traversed by moving objects, such as vehicles, where each path consists of a sequence of connected road segments. These representations are essential for various downstream tasks, including travel-time estimation and path ranking. However, existing methods often suffer from ineffective view generation and inefficient training procedures due to their reliance on heuristic augmentations and computationally intensive contrastive learning. To address these issues, we propose AutoPRL, an effective and efficient self-supervised PRL framework that integrates three key components. The Automatic Data Augmentation (AutoDA) module employs variational autoencoders to generate semantically meaningful and diverse feature-level path views by modeling each topology-aware node embedding with a Gaussian latent distribution, thereby mitigating information loss and enhancing representation quality. Second, the Cross-View Decoupled Contrastive (CVDC) learning mechanism improves training efficiency by disentangling the optimization of positive and negative pairs across node- and path-level views, enabling more stable contrastive learning. Third, an Adaptive Co-Training (ACT) module dynamically adjusts uncertainty-based loss weights to balance multiple objectives, ensuring efficient and scalable training. Extensive experiments on three real-world datasets demonstrate that AutoPRL achieves strong and consistent performance in both effectiveness and efficiency across multiple downstream tasks.

Figures & tables

Fig. 1: Architectures of the conventional contrastive paradigm and our AutoPRL for path representation learning. In (a), heuristic augmentation directly perturbs the discrete path sequence. In (b), AutoDA generates feature-level views as continuous embedding sequences from topology-aware node embeddings, rather than discrete node sequences on the road-network graph. Contrast denotes the contrastive loss.

Fig. 2: An example of heuristic path view generation.

Fig. 3: Solution Overview.

Fig. 4: Overall architecture of AutoPRL . The framework consists of three modules: AutoDA for automatic path view generation, CVDC for contrastive representation learning with cross-path and cross-node objectives, and ACT for adaptive multi-objective optimization.

Fig. 5: Automatic path view generation in AutoDA . The input path is first mapped into topology-aware spatial embeddings. Two VNAEs then encode the path into hidden representations, from which MLPs parameterize Gaussian latent distributions. By sampling from these distributions, AutoDA generates two semantically consistent feature-level views, $p_{i}^{(1)}$ and $p_{i}^{(2)}$ .

Fig. 6: Illustration of the progress-based prior. The approximate posterior is aligned with a dynamic Gaussian prior through KL regularization. During training, the mean of the prior gradually shifts from $\boldsymbol{\mu}_{s_{1}}$ to $\boldsymbol{\mu}_{s_{T}}$ , which improves the alignment between the prior and the evolving latent space.

Method		Aalborg			Chengdu			Harbin
Method		MAE $\downarrow$	MAPE (%) $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	MAPE (%) $\downarrow$	RMSE $\downarrow$	MAE $\downarrow$	MAPE (%) $\downarrow$	RMSE $\downarrow$
Supervised	PathRank	2.913	18.713	4.832	5.802	34.014	8.894	4.127	14.339	5.301
	HierETA	3.097	18.620	5.058	5.060	28.820	7.921	4.897	16.506	5.961
	CompactETA	3.030	19.726	4.791	4.832	28.678	7.447	4.372	14.974	5.362
	HMTRL	3.515	19.955	5.792	4.861	28.529	7.473	4.704	15.543	5.803
Unsupervised	Node2vec	5.701	35.485	8.165	5.997	35.919	8.653	5.399	18.709	6.773

TABLE I: Overall performance on travel time estimation. The best results are highlighted in bold, and the second-best results are underlined.

Method		Aalborg			Chengdu			Harbin
Method		MAE $\downarrow$	TAU ( $\tau$ ) $\uparrow$	RHO ( $\rho$ ) $\uparrow$	MAE $\downarrow$	TAU ( $\tau$ ) $\uparrow$	RHO ( $\rho$ ) $\uparrow$	MAE $\downarrow$	TAU ( $\tau$ ) $\uparrow$	RHO ( $\rho$ ) $\uparrow$
Supervised	PathRank	0.159	0.727	0.751	0.123	0.703	0.742	0.145	0.821	0.832
	HierETA	0.152	0.757	0.767	0.178	0.735	0.774	0.133	0.846	0.879
	CompactETA	0.169	0.698	0.709	0.195	0.661	0.705	0.160	0.796	0.833
	HMTRL	0.163	0.729	0.741	0.149	0.779	0.789	0.133	0.838	0.874
Unsupervised	Node2vec	0.253	0.632	0.637	0.246	0.464	0.505	0.211	0.468	0.512

TABLE II: Overall performance on path ranking. The best results are highlighted in bold, and the second-best results are underlined.

Fig. 7: Effect of Pre-training.

Method	Chengdu			Harbin
Method	10K	30K	50K	60K	80K	100K
Toast	5.233	7.254	9.283	7.674	9.231	11.643
PIM	4.107	6.788	8.328	5.962	7.479	9.896
START	2.765	4.371	5.287	4.102	5.997	8.122
LightPath	2.431	3.289	5.364	3.562	5.296	7.432
RED	1.453	1.578	1.690	2.590	3.102	3.457

TABLE III: Evaluation on Most Similar Path Search Task.

Method	Aalborg			Chengdu
Method	Acc.	Pre.	F1	Acc.	Pre.	F1
Toast	0.692	0.711	0.738	0.774	0.785	0.806
PIM	0.722	0.739	0.762	0.785	0.799	0.813
START	0.751	0.770	0.787	0.804	0.826	0.853
LightPath	0.752	0.768	0.790	0.817	0.832	0.849
RED	0.785	0.811	0.830	0.844	0.860	0.879

TABLE IV: Evaluation on Path Recommendation Task.

Method	Aalborg			Chengdu
Method	R@3	R@5	MAP	R@3	R@5	MAP
Toast	0.656	0.705	0.697	0.770	0.797	0.800
PIM	0.679	0.716	0.663	0.792	0.837	0.845
START	0.708	0.725	0.692	0.834	0.860	0.863
LightPath	0.714	0.756	0.736	0.830	0.855	0.823
RED	0.733	0.778	0.746	0.862	0.881	0.863

TABLE V: Evaluation on Path Imputation Task.

Fig. 8: Ablation Study.

Fig. 9: Effect of Temperature

t

Fig. 10: Scalability Performance vs. Cardinality

Fig. 11: (a) Model training time (in minutes) per epoch; (b) model inference time (in ms) per path; and (c) – (e) training convergence of AutoPRL on three datasets.

Method	Travel Time Estimation						Path Ranking
	Aalborg			Chengdu			Aalborg			Chengdu
	MAE	MAPE	RMSE	MAE	MAPE	RMSE	MAE	TAU ( $\tau$ )	RHO ( $\rho$ )	MAE	TAU ( $\tau$ )	RHO ( $\rho$ )
AutoPRL-DGM	1.976	12.901	3.396	4.683	27.447	7.091	0.079	0.870	0.878	0.112	0.821	0.830
AutoPRL-GMGP	1.971	12.867	3.382	4.672	27.318	7.074	0.073	0.890	0.893	0.103	0.836	0.845

TABLE VI: Ablation study on travel-time estimation and path ranking under different generative modeling paradigms.

Training Epochs	Aalborg								Harbin
	AutoPRL-SimCLR				AutoPRL-DCLW				AutoPRL-SimCLR				AutoPRL-DCLW
	16	32	64	128	16	32	64	128	16	32	64	128	16	32	64	128
400	2.844	2.657	2.539	2.340	2.112	2.044	1.971	1.954	–	–	–	–	–	–	–	–
500	2.435	2.216	2.161	2.064	–	–	–	–	6.158	5.725	5.384	5.008	4.001	3.964	3.883	3.775
600	–	–	–	–	–	–	–	–	5.574	4.977	4.472	4.258	–	–	–	–
700	–	–	–	–	–	–	–	–	4.476	4.433	4.212	3.979	–	–	–	–

TABLE VII: Batch-size Sensitivity Analysis on Travel-Time Estimation Comparing AutoPRL-SimCLR and AutoPRL-DCLW .

Method	Travel Time Estimation						Path Ranking
	Aalborg			Chengdu			Aalborg			Chengdu
	MAE	MAPE	RMSE	MAE	MAPE	RMSE	MAE	TAU ( $\tau$ )	RHO ( $\rho$ )	MAE	TAU ( $\tau$ )	RHO ( $\rho$ )
Toast	3.439	21.311	5.555	4.918	28.567	7.513	0.138	0.768	0.779	0.131	0.734	0.752
Toast w/ noise	4.426	21.906	6.066	5.622	29.134	8.233	0.207	0.672	0.683	0.206	0.692	0.703
PIM	3.471	21.314	5.640	4.924	28.553	7.493	0.141	0.764	0.777	0.125	0.725	0.745
PIM w/ noise	4.003	21.904	5.982	5.427	28.994	7.89	0.192	0.727	0.738	0.169	0.674	0.685

TABLE VIII: Performance on Travel-Time Estimation and Path Ranking under Noisy Representations.

Fig. 13: Case study of path view generation. We compare the original path, a random dropping view, and an AutoDA -generated view that is projected onto the road network only for visualization. Random dropping may remove important intermediate nodes and break local path continuity, while AutoDA better preserves the overall route semantics.

Representation Learning

Representation learning develops data transformations that expose useful structure for prediction, transfer, generation, and control.

See all 813 papers

DGCPath: Distribution-Aware Generative Contrastive Framework for Self-supervised Path Representation Learning -- Extended Version
Sep 7, 2026Sean Bin Yang, Hao Miao, Zongyi Xu +5Contrastive Learning Representation Learning
Aalborg University, Denmark · The Hong Kong Polytechnic University, Hong Kong · Chongqing Univeristy of Posts and Telecommunications, China +1
TraveL: Transformer-based Multi-view Path Distributional Representation Learning
Sep 3, 2026Fang He, Tao-yang Fu, Wang-chien LeeRepresentation Learning Graph Representation Learning
The Pennsylvania State University
REFINE: Trajectory Representation Learning via Closed-Loop Transcription -- Extended Version
Sep 7, 2026Sean Bin Yang, Ying Sun, Jilin Hu +5Contrastive Learning Trajectory Reconstruction
Aalborg University Aalborg, Denmark · Chongqing Universtiy of Posts and Telecommunications Chongqing, China · East China Normal University Shanghai, China +1

Sep 7, 2026J/K move · Enter open · S save

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

Sean Bin Yang, Hao Miao, Zongyi Xu, Jilin Hu+4

Aalborg University, Denmark · The Hong Kong Polytechnic University, Hong Kong · Chongqing Univeristy of Posts and Telecommunications, China+1

Contrastive Learning Representation Learning Self-Supervised Time Series Representation Learning Trajectory Representation Learning

Open paper AskPDF arXiv

Due to the proliferation of vehicle trajectory data enabled by advanced sensing technologies, path representation learning has become a pivotal task in intelligent transportation systems. Although existing self-supervised approaches have achieved promising performance, their dependence on deterministic contrastive learning paradigms and handcrafted view augmentation strategies inherently restricts their cross-scenario generalization capabilities. To address these limitations, we present DGCPath, an innovative Distribution-aware Generative Contrastive learning framework for Path representation. This framework establishes a synergistic connection between generative modeling and distributional contrastive learning, enabling the acquisition of robust and transferable feature embeddings. Specifically, our framework incorporates: (1) a diffusion-based view generator that autonomously produces semantically coherent yet diverse trajectory views from Gaussian noise; (2) a variational contrastive mechanism that enforces latent feature alignment at the distribution level, transcending conventional instance-wise consistency; and (3) a novel generative cross-supervision module that reinforces view-level consistency through cross-view reconstruction learning. Comprehensive evaluations on three real-world trajectory datasets demonstrate that DGCPath outperforms state-of-the-art baselines on two distinct downstream tasks, validating its enhanced generalization capability and representation effectiveness.

Abstract

Representation Learning

Explore similar work

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

TraveL: Transformer-based Multi-view Path Distributional Representation Learning

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