Quantifying travel sequence predictability via a constrained vector-quantized variational autoencoder

Zhi Li

doi:10.1016/j.trc.2026.105958

Quantifying travel sequence predictability via a constrained vector-quantized variational autoencoder

2026-08-22

0

PRE

AI

DOI：10.1016/j.trc.2026.105958

Original

Original request for help

Save

Abstract

En 中文

• Extend predictability from discrete states to continuous travel sequences. • A constrained VQ-VAE computes the predictability–distortion curve. • Validate on synthetic data against a theoretical predictability-distortion curve. • Real-world mobility data validation quantifies cross-vehicle predictability heterogeneity. Abstract A travel sequence is a time-ordered record of travel-related states (e.g., location, speed, and energy consumption rate). Understanding the intrinsic predictability of such sequences is useful for characterizing behavioral regularity and benchmarking predictive difficulty in mobility data. While classical entropy-based methods have been effective for analyzing discrete symbolic sequences, real-world mobility data such as GPS trajectories, driving speed, and battery state are inherently continuous and often form complex, multi-dimensional time series. This disconnect limits the applicability of existing predictability measures in practical transportation settings. In this study, we develop an information-theoretic representation-learning framework for extending entropy-based predictability analysis from discrete sequences to continuous behavioral signals under an explicit distortion constraint. By combining a Vector-Quantized Variational Autoencoder with augmented Lagrangian optimization, the framework enables principled predictability estimation while accounting for discretization-induced information loss. Experiments on synthetic data and an open-access San Francisco taxi benchmark confirm both the theoretical soundness of the approach and its practical ability to recover meaningful predictability patterns from real-world trajectory data. Large-scale experiments using 16 months of operational data from 3777 electric vehicles in Shanghai further demonstrate the framework’s effectiveness. Under the D4 setting of a 5% state-of-charge error and 2.5 km spatial tolerance, 89.70% of vehicles satisfy the realized inference-time distortion constraint; among these theoretically admissible cases, the mean behavioral predictability is 94.60%. The implementation code and reproducible experiments are publicly available at https://github.com/zhilee2023/Quantifying-the-Predictability-of-Travel-Sequences

Keywords:

Predictability

VQ-VAE

Electric vehicle

Information-theoretic metrics

Quantifying travel sequence predictability via a constrained vector-quantized variational autoencoder

Abstract

Journal

Organization

Cited Papers