PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Sean Bin Yang, Hao Miao, Zongyi Xu, Jilin Hu, Xiangmeng Wang, Hua Lu, Bin Yang, Christian S. Jensen

arXiv:2609.07316v1cs.AI

TL;DR

Path representation learning is constrained by deterministic contrastive objectives and handcrafted view augmentations, motivating a more transferable approach. DGCPath combines diffusion-generated views, distribution-level contrastive alignment, and generative cross-supervision, outperforming state-of-the-art methods across three datasets and two downstream tasks.

  • Problem

    Existing self-supervised path representation methods rely on deterministic contrastive learning and handcrafted view augmentations, limiting cross-scenario generalization.

    Source: §1 · p. 1

  • Method

    DGCPath combines a diffusion-based view generator, distribution-aware variational contrastive learning, and generative cross-supervision for path representation learning.

    Source: §1 · p. 2, §6 · p. 14

  • Results

    DGCPath consistently outperforms state-of-the-art methods across three real-world datasets and two downstream tasks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The framework yields more generalizable path representations and improved downstream-task performance.

    Source: §6 · p. 14

    Self-reconstruction may allow the encoder to preserve view-specific or coordinate-dependent information.

    Source: §4.4 · p. 9

Abstract

from arXiv · show

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.

1 Introduction

Path representation learning is important for intelligent transportation analysis, but deterministic contrastive objectives and handcrafted augmentations limit representation flexibility and transferability. DGCPATH addresses these constraints by combining diffusion-based view generation with distributional and generative contrastive learning, and achieves consistent gains across datasets and tasks.

  • Vehicle path representations encode raw trajectories as compact vectors that support traffic analysis, path recommendation, and traffic prediction.Source: §1 · p. 1
  • Existing methods rely on deterministic contrastive samples and pre-defined augmentations, which restrict representation expressiveness, adaptability, and cross-task transfer.Source: §1 · p. 2
  • DGCPATH models path representations as latent probability distributions and aligns distributions rather than fixed instances to learn smoother, uncertainty-aware embeddings.Source: §1 · p. 2
  • Its diffusion-based view generator produces diverse, semantically consistent path views from noise without manually designed augmentation heuristics.Source: §1 · p. 2
  • The framework combines distributional contrastive learning with generative cross-supervision to align views of the same path and separate different paths.Source: §1 · p. 2
  • DGCPATH consistently outperforms state-of-the-art methods across three real-world datasets and two downstream tasks in representation quality and downstream performance.Source: §1 · p. 2

2 Related Work

Prior path representation learning work primarily studies single-modality path sequences and develops embeddings for downstream intelligent transportation applications.

  • Existing path representation methods mostly focus on single-modality path-sequence data.Source: §2 · p. 2

3 Preliminaries

The preliminaries define road networks as directed graphs and paths as topologically continuous ordered node sequences, then formulate representation learning as mapping paths to latent vectors.

  • A road network is modeled as a directed graph G = (V, E), with nodes representing intersections and directed edges representing road segments.Source: §3 · p. 2
  • A path is an ordered sequence of connected nodes whose consecutive nodes are linked by directed edges.Source: §3 · p. 3
  • The objective is to learn a function fθ that maps each path to a d-dimensional representation vector in latent space.Source: §3 · p. 3

4 Methodology

DGCPath combines distributional contrastive learning with generative cross-supervision to learn path representations that are view-invariant, path-discriminative, and semantically informative. Its theoretical analysis explains how the objectives suppress view-specific information, separate different paths, and preserve shared semantics.

Distributional Contrastive Learning

  • DGCPath aligns latent distributions of positive path views while separating distributions from different paths.The objective uses small JS divergence for positive pairs and large JS divergence for negative pairs.Source: §4.3 · p. 4, §4.4 · p. 8

Invariant Feature Learning via Generative Cross-Supervision

  • Generative cross-supervision encourages reconstructions from one view to preserve semantic information shared with another view.Its trainable reconstruction component can be reduced by improving representation of the shared semantic component S.Source: §4.3 · p. 7, §4.5 · p. 10
  • Unlike self-reconstruction, cross-supervision removes the shortcut of exploiting view-specific nuisance factors.The nuisance factor in one view is not predictive of the nuisance factor in the other view, discouraging augmentation-specific or coordinate-dependent information.Source: §4.3 · p. 7, §4.4 · p. 9

Overall Theoretical Interpretation

  • Jointly optimizing the proposed objectives yields representations that are simultaneously view-invariant, path-discriminative, and semantically informative.The theoretical interpretation assigns complementary roles to distributional alignment, distributional contrast, and generative cross-supervision.Source: §4.3 · p. 7, §4.5 · p. 10

View Invariance via Distributional Alignment

  • Minimizing positive-pair JS divergence suppresses augmentation-specific information and reduces prediction discrepancy across views.The analysis connects distributional alignment to conditional view invariance and an upper bound on view-dependent prediction differences.Source: §4.3 · p. 6, §4.5 · p. 9

Path Discriminability via Distributional Contrast

  • The negative contrastive term enlarges the margin between positive and negative path pairs, improving latent-distribution discriminability.As the margin increases, the corresponding prediction-error upper bound decreases exponentially.Source: §4.3 · p. 6, §4.5 · p. 10

5 Experiments

Experiments evaluate path representations on three real-world datasets across travel time estimation and path ranking, using unsupervised pre-training followed by supervised regression.

Datasets

  • Three datasets—Aalborg, Chengdu, and Harbin—provide road networks and GPS trajectories for evaluation.The road networks contain 10,017, 6,632, and 8,497 nodes, respectively.Source: §5 · p. 11

Downstream Tasks

  • The study evaluates travel time estimation and path ranking as downstream tasks.Travel time estimation uses MAE, MAPE, and RMSE; path ranking assigns preference scores to candidate paths.Source: §5 · p. 11

Baselines

  • Ten baselines span six unsupervised and four supervised path-representation methods.Several unsupervised baselines rely on predefined augmentations such as random masking or node dropping.Source: §5 · p. 11

Implementation

  • All baseline representations are set to 128 dimensions, with optimization conducted using AdamW and a cosine-decay schedule over 600 epochs.The schedule includes a 50-epoch warm-up period.Source: §5 · p. 11

5.1 Performance Findings

DGCPath achieves the strongest reported downstream performance across all three datasets and both tasks, with diffusion-based generation and distribution-aware contrastive learning identified as contributing factors.

Overall Results

  • DGCPath achieves the best performance across all baselines on both downstream tasks and all three datasets.The evaluation covers travel time estimation and path ranking.Source: §5.1 · p. 11

Travel Time Estimation

  • 13.27% and 2.21% are the reported maximum reductions in Aalborg travel-time MAE and RMSE, respectively.These figures concern DGCPath’s comparison with the evaluated baselines.Source: §5.1 · p. 11

Path Ranking

  • 4.66% and 6.94% are the reported maximum increases in Harbin Kendall’s τ and Spearman’s ρ, respectively.These figures concern the path-ranking task.Source: §5.1 · p. 11

Interpretation

  • DGCPath’s performance is attributed to integrating a diffusion-based view generator with distribution-aware contrastive learning.The combination is described as capturing spatial structure and representation diversity.Source: §5.1 · p. 12

5.2 Ablation Study

Ablation results indicate that all three major components contribute to DGCPath, with the variational contrastive module having the largest reported impact on generalizable representations.

Component Contributions

  • Removing any of DVG, VCM, or GCSM causes performance to drop on both downstream tasks.The ablations evaluate the diffusion-based view generator, variational contrastive module, and generative cross-supervision module.Source: §5.2 · p. 12
  • The VCM has a more critical role than DVG and GCSM in learning generalizable path representations.The reported impact varies by downstream task.Source: §5.2 · p. 12
  • DVG and GCSM provide complementary contributions to the overall model.Removing either component also degrades performance.Source: §5.2 · p. 12

Pre-training Effects

  • With DGCPath pre-training, PathRank matches the 12K-label from-scratch performance using 8K, 9K, and 9K labels on Aalborg, Chengdu, and Harbin.These figures are reported for travel time estimation.Source: §5.3 · p. 12
  • PathRank with pre-training outperforms its non-pretrained counterpart when both use the full 12K labeled samples.The comparison is reported after transferring DGCPath’s learned encoder parameters.Source: §5.3 · p. 12

5.4 Path Recommendation

The path recommendation evaluation extends DGCPATH testing to recommend user-chosen paths among alternatives, while parameter studies examine temperature, embedding dimensionality, and diffusion depth.

Path Recommendation

  • Path recommendation labels the user-taken trajectory as 1 and alternative paths with the same origin and destination as 0.Performance is evaluated using Accuracy, Precision, and F1-score, with higher values indicating better recommendation quality.Source: §5.4 · p. 13

Temperature

  • The parameter study finds the best performance at temperature t = 0.5.A moderately warm temperature balances contrastive-objective sharpness, mitigating overly confident similarity assignments and improving optimization stability and representation quality.Source: §5.5 · p. 13

Embedding Dimensionality

  • Embedding dimensionality performs best at d = 128, providing a favorable balance between representation capacity and overfitting.Source: §5.5 · p. 13

Diffusion Step

  • DGCPATH achieves the best performance when the diffusion step is T = 1000.Smaller diffusion depths may provide insufficient perturbation, whereas excessively large depths can introduce unnecessary noise that degrades representation quality.Source: §5.5 · pp. 13–14

5.6 Effect of Dynamic Traffic State

The dynamic traffic state experiment adds temporal traffic information to DGCPATH and reports further performance improvement.

Dynamic Traffic State

  • Incorporating dynamic traffic states further improves DGCPATH performance.The variant DGCPATH w/ DTS explicitly considers temporal traffic information constructed from time-varying traffic-flow sequences.Source: §5.6 · p. 14

5.7 Model Scalability

The scalability analysis evaluates DGCPATH with different training-data proportions and finds consistently better performance as more training data is used.

Training Data Size

  • DGCPATH shows consistent performance improvement as the training-data proportion increases from 20% to 100%.The analysis uses the Aalborg dataset and compares training proportions of 20%, 40%, 60%, 80%, and 100%.Source: §5.7 · p. 14

5.8 Robustness Analysis

The robustness analysis evaluates noisy path representations and reports that DGCPATH remains robust while outperforming comparison methods.

Noisy Representations

  • DGCPATH consistently outperforms comparison methods under noisy representations through a smaller performance drop percentage.Gaussian perturbations, including σ = 0.03, are injected into path representations before training a regressor on the Aalborg and Chengdu datasets.Source: §5.8 · p. 14

6 Conclusion

DGCPATH combines learnable view generation, distribution-aware contrastive learning, and generative cross-supervision to improve the generalizability of path representations across downstream tasks.

  • DGCPATH generates two path views with a learnable diffusion-based generator, aligns positive-pair distributions while separating positive and negative pairs, and applies generative cross-supervision.Source: §6 · p. 14

Related papers

Loading 2609.07316v1…