TTCD: Transformer Integrated Temporal Causal Discovery from Non-Stationary Time Series Data

Researchers have introduced the Transformer Integrated Temporal Causal Discovery (TTCD) Framework, a novel end-to-end approach designed to identify causal relationships in complex, non-stationary time series data. Addressing limitations in existing constraint-based and score-based methods, which often struggle with noisy, nonlinear settings or rely on restrictive statistical assumptions, TTCD integrates temporal and frequency-domain attention mechanisms. The framework features a Non-Stationary Feature Learner and a custom Causal Structure Learner. A key innovation is reconstruction-guided causal signal distillation, which uses the transformer decoder to filter noise and spurious correlations while preserving meaningful dependencies. This allows for the inference of underlying causal graphs without strict assumptions about noise distributions. Experimental results across synthetic, benchmark, and real-world datasets demonstrate that TTCD consistently outperforms state-of-the-art baselines in accuracy and alignment with domain knowledge. This advancement offers significant potential for fields such as environmental science, epidemiology, and economics, where robust causal discovery in dynamic environments is critical for accurate modeling and decision-making.