TTCD: Transformer-Based Causal Discovery for Non-Stationary Time Series
A new framework, Transformer Integrated Temporal Causal Discovery (TTCD), has been proposed for learning causal relations from non-stationary time series data. The method addresses limitations of existing constraint-based and score-based approaches, which struggle with limited data, complex distributions, and non-stationarity. TTCD introduces a Non-Stationary Feature Learner that integrates temporal and frequency-domain attention mechanisms. The framework is designed for applications in environmental science, epidemiology, and economics, where robust causal discovery is critical. The paper is available on arXiv.
Key facts
- TTCD stands for Transformer Integrated Temporal Causal Discovery.
- The framework is designed for non-stationary time series data.
- It uses a Non-Stationary Feature Learner with temporal and frequency-domain attention.
- Existing methods rely on conditional independence tests or strong statistical assumptions.
- TTCD is an end-to-end approach for learning contemporaneous and lagged causal relations.
- The paper is published on arXiv with ID 2605.08111.
- Applications include environmental science, epidemiology, and economics.
- The method addresses noisy and nonlinear settings.
Entities
Institutions
- arXiv