Traffic time-series forecasting faces significant challenges from varying data complexity and domain-specific temporal patterns that existing transformer approaches fail to address through fixed architectural configurations. This article introduces the adaptive complexity-aware traffic transformer (ACTFormer), a novel framework that establishes adaptive processing strategies based on data characteristics. ACTFormer contributes four key innovations: entropy-based complexity analysis for traffic pattern quantification, differentiable adaptive vocabulary selection via Gumbel-Softmax relaxation, enabling end-to-end optimization, traffic-aware contextual encoding capturing domain-specific dependencies, and comprehensive experimental validation demonstrating effectiveness across diverse scenarios. The adaptive mechanism intelligently matches computational resources to problem difficulty, with high-complexity scenarios benefiting from large vocabularies (1024 tokens), achieving 15.6% performance gains. Comprehensive experiments across six benchmark tasks (PeMS04, PeMS07, PeMS08, METR-LA, NYCTaxi Drop-off, and NYCTaxi Pick-up) demonstrate superior performance against 34 baseline methods; on the PeMS and NYCTaxi tasks, ACTFormer achieves consistent 8.7%-14.6% mean absolute error (MAE) improvements over the strongest transformer (STGAFormer) baseline, with strong correlation (0.84) between complexity scores and selected vocabulary sizes. ACTFormer's computational efficiency (1.181 M parameters and 29.49 s training time) enables immediate deployment in intelligent transportation systems (ITSs) while establishing adaptive complexity-aware processing as a fundamental principle for transformer architectures.
Yang et al. (Thu,) studied this question.