Revisiting the Modifiable Areal Unit Problem in Deep Traffic Prediction with Visual Analytics.

Area of Science:

• Urban planning and transportation science
• Computer science and artificial intelligence
• Geographic information systems and spatial analysis

Background:

• Deep learning models for urban traffic prediction rely on aggregated spatiotemporal data, which can be sensitive to the Modifiable Areal Unit Problem (MAUP).
• MAUP can destabilize neural network inputs, feature embeddings, and predictions, limiting the utility of deep learning for traffic forecasting experts.
• Existing methods lack effective tools to visualize and analyze the impact of data aggregation scales on deep learning model performance.

Purpose of the Study:

• To address the MAUP in urban traffic prediction by developing a visual analytics solution for investigating multi-scalar data aggregations.
• To enable experts to understand and mitigate the impact of varying spatial scales on deep learning model inputs and prediction errors.
• To facilitate the development and refinement of deep traffic prediction models through interactive exploration.

Main Methods:

• Leveraging unit visualization techniques to explore many-to-many relationships between multi-scalar traffic data aggregations and neural network predictions.
• Developing a visual analytics solution integrating a Bivariate Map for spatial error depiction, a Moran's I Scatterplot for spatial association analysis, and a Multi-scale Attribution View for model comparison.
• Conducting case studies with real-world Shenzhen taxi trip data and expert interviews for evaluation.

Main Results:

• Geographical scale variations significantly impact the performance of deep traffic prediction models.
• Interactive visual exploration of dynamically varying inputs and outputs aids experts in developing more robust deep traffic prediction models.
• The proposed visual analytics solution effectively visualizes input traffic data, prediction errors, and model attributions across different scales.

Conclusions:

• Visual analytics is crucial for understanding and addressing the MAUP in deep learning-based urban traffic prediction.
• The developed integrated visualization system empowers domain experts to analyze scale-dependent prediction performance and refine deep learning models.
• This approach enhances the reliability and interpretability of deep learning models in urban traffic forecasting.