ACTFormer: Adaptive Complexity-Aware Traffic Transformer for Intelligent Flow Prediction.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Traffic time-series forecasting is challenged by complex data and unique temporal patterns.
• Existing transformer models use fixed architectures, failing to adapt to varying data characteristics.

Purpose of the Study:

• Introduce the Adaptive Complexity-Aware Traffic Transformer (ACTFormer) framework.
• Enable adaptive processing strategies tailored to specific traffic data characteristics.
• Enhance the accuracy and efficiency of traffic time-series forecasting.

Main Methods:

• Developed entropy-based complexity analysis for traffic pattern quantification.
• Implemented differentiable adaptive vocabulary selection using Gumbel-Softmax relaxation for end-to-end optimization.
• Incorporated traffic-aware contextual encoding to capture domain-specific dependencies.

Main Results:

• ACTFormer achieved performance gains of 15.6% in high-complexity scenarios by utilizing larger vocabularies (1024 tokens).
• Demonstrated superior performance across six benchmark datasets (PeMS and NYCTaxi), outperforming 34 baseline methods.
• Achieved 8.7%-14.6% Mean Absolute Error (MAE) improvements over the STGAFormer baseline, with a 0.84 correlation between complexity scores and vocabulary size.

Conclusions:

• ACTFormer establishes adaptive complexity-aware processing as a fundamental principle for transformer architectures.
• The framework's computational efficiency allows for immediate deployment in intelligent transportation systems (ITSs).
• ACTFormer offers a novel and effective approach to handling the complexities of traffic time-series forecasting.