[Paper Review] Transformer-based Map Matching Model with Limited Ground-Truth Data using Transfer-Learning Approach

Zhixiong Jin, Kim, Jiwon+2 more|arXiv (Cornell University)|Aug 1, 2021

Traffic Prediction and Management Techniques20 references4 citations

TL;DR

This paper proposes a Transformer-based map-matching model that leverages transfer learning to overcome data scarcity, pre-training on synthetically generated trajectories and fine-tuning with limited real ground-truth data. The model achieves state-of-the-art performance on the Gangnam taxi dataset, outperforming rule-based and deep learning baselines, while attention analysis reveals meaningful internal and external data correlations.

ABSTRACT

In many spatial trajectory-based applications, it is necessary to map raw trajectory data points onto road networks in digital maps, which is commonly referred to as a map-matching process. While most previous map-matching methods have focused on using rule-based algorithms to deal with the map-matching problems, in this paper, we consider the map-matching task from the data-driven perspective, proposing a deep learning-based map-matching model. We build a Transformer-based map-matching model with a transfer learning approach. We generate trajectory data to pre-train the Transformer model and then fine-tune the model with a limited number of ground-truth data to minimize the model development cost and reduce the real-to-virtual gap. Three metrics (Average Hamming Distance, F-score, and BLEU) at two levels (point and segment level) are used to evaluate the model performance. The results indicate that the proposed model outperforms existing models. Furthermore, we use the attention weights of the Transformer to plot the map-matching process and find how the model matches the road segments correctly.

Motivation & Objective

• Address the challenge of limited real-world ground-truth trajectory data for training high-performing map-matching models.
• Improve map-matching accuracy by leveraging collective mobility patterns and noise characteristics from large-scale historical trajectory data.
• Develop a data-driven deep learning approach that captures complex spatio-temporal correlations in GPS trajectories, overcoming limitations of rule-based and independent trajectory processing methods.
• Enable effective transfer learning from synthetic to real-world data to reduce the real-to-virtual performance gap in map-matching tasks.
• Provide interpretability of model decisions through attention mechanism analysis, revealing internal GPS point correlations and external relationships with road segments.

Proposed method

• Pre-train a Transformer-based map-matching model on a large volume of synthetically generated GPS trajectories derived from road network topology.
• Fine-tune the pre-trained model using a limited number of real ground-truth trajectory datasets to minimize domain shift and improve real-world performance.
• Employ a sequence-to-sequence architecture with multi-head self-attention to model long-range dependencies and contextual relationships among GPS points.
• Use data augmentation via road network-based trajectory generation to simulate realistic mobility patterns and noise characteristics.
• Apply transfer learning to transfer knowledge from synthetic data (pre-training) to real data (fine-tuning), reducing reliance on expensive real annotations.
• Evaluate model performance using point-level and segment-level metrics: Average Hamming Distance (AHD), F-score, and BLEU, ensuring comprehensive assessment of matching accuracy.

Experimental results

Research questions

• RQ1 Can a Transformer-based model pre-trained on synthetic trajectories and fine-tuned on limited real ground-truth data achieve superior map-matching performance compared to existing rule-based and deep learning models?
• RQ2 How does the attention mechanism in the Transformer model reveal internal correlations among GPS points and external relationships between input trajectories and matched road segments?
• RQ3 To what extent does transfer learning from synthetic to real data reduce the real-to-virtual performance gap in map-matching tasks?
• RQ4 How do different data generation strategies for pre-training affect the final model performance after fine-tuning?
• RQ5 Can the model generalize well to real-world urban environments despite limited real ground-truth data, and how does it compare to baselines in terms of accuracy and robustness?

Key findings

• The pre-trained Transformer model outperformed rule-based models (FMM and ST-matching) on the pre-training stage, demonstrating the effectiveness of synthetic data for initial representation learning.
• The fine-tuned Transformer model achieved state-of-the-art performance across three datasets, surpassing both LSTM-based seq2seq and attentional seq2seq baselines in all evaluation metrics.
• Fine-tuning significantly reduced the real-to-virtual gap: in the better-performing pre-trained model, fine-tuning was essential for realistic matching; in the weaker model, it primarily improved performance.
• For a representative noisy trajectory, the fine-tuned model correctly matched the first two GPS points to their corresponding road segments, illustrating improved robustness.
• Attention analysis revealed strong correlations between target GPS points and points on the current and adjacent road segments, confirming the model's ability to capture spatial context.
• The model demonstrated that points on neighboring segments also contribute to matching decisions, indicating effective modeling of spatial relationships beyond the immediate segment.