Causal Temporal Neural Networks Using the Conditional Average Treatment Effect

This paper presents a method to integrate causal inference into deep learning for time series forecasting. We consider time series for complex systems characterized by non-linear dynamics, high dimensionality, and uncertainty. The challenge of effectively capturing temporal dependencies persists due to the prevalence of spurious correlations. To overcome this, our method integrates Long Short-Term Memory (LSTM) for sequence forecasting with prior knowledge about the causal structure of the system. For this, we introduce a causal regularization term that controls for the Conditional Average Treatment Effect (CATE). Experimental results across real and synthetic datasets demonstrate superior performance compared to state-of-the-art models. Furthermore, an ablation study highlights the critical role of causal regularization graph-based interventions alongside causal feature selection. By embedding a learned causal graph derived from causal discovery to identify the optimal predictors that improve model performance and reduce uncertainties.