Long-sequence model for traffic forecasting in suboptimal situation

Accurate traffic prediction generally depends on reliable and noise-free data, which may not reflect real-world scenarios. Additionally, the high sampling rate poses challenges in capturing the inherent temporal periodicity present in traffic data. To address these issues and achieve long-term traffic prediction, this paper proposes a novel approach called as Spatio-Temporal State Space Probabilistic Diffusion (ST-SSPD). The proposed method utilizes state space modeling to decompose the historical signal of an ego-graph into the frequency domain, thereby reducing the impact of missing data during the subsequent reconstruction phase. The reconstructed signal is then employed to condition a probabilistic denoising diffusion model, enabling the learning of the underlying data distribution by the gradual denoising of pure Gaussian noise into the desired prediction. This technique facilitates the modeling of long-range dependencies by representing an optimal historical approximation through a high-order polynomial projection, resulting in significantly enhanced traffic prediction accuracy. Extensive experiments were conducted on real-world traffic datasets with varying levels of data loss to evaluate the effectiveness of the proposed model. The results demonstrate the model's robustness to data loss and its superior performance compared to existing state-of-the-art architectures in traffic forecasting.