Lightweight and Interpretable Transformer via Mixed Graph Algorithm Unrolling for Traffic Forecast

The article presents a novel lightweight and interpretable transformer-like neural network for traffic forecasting, utilizing mixed-graph algorithm unrolling. It constructs two graphs—an undirected graph for spatial correlations and a directed graph for temporal relationships—and employs an iterative algorithm based on the alternating direction method of multipliers (ADMM) to enhance data-driven parameter learning. This approach not only maintains competitive forecasting performance compared to state-of-the-art methods but also significantly reduces the parameter count, making it advantageous for practitioners seeking efficient and interpretable models in AI applications.