Temporal Causal Discovery Based on Multi-scale Attention Fusion

Temporal causal discovery is essential in data mining and machine learning for uncovering causal relationships in time series data. Traditional methods struggle with complex, multi-scale periodicity and fail to capture dynamic dependencies across various time scales. We propose a multi-scale attention fusion approach to address these challenges, integrating local and global dependencies via a dual-branch network. The local branch uses an enhanced Temporal Convolutional Network with FFT-based period detection for adaptive multi-scale patching, while the global branch employs an iTransformer to capture long-range dependencies. Attention scores are fused using learnable weights, and causality is validated with an Enhanced Permutation Importance Test (EPIT). Our framework excels in identifying complex causal structures in non-linear, high-dimensional time series, significantly enhancing accuracy and granularity of causal insights.