Abstract

This paper presents a causal discovery model for time series analysis, which can find several causal lags of one variable over another, presenting the results by means of a temporal causal graph. The proposed model is based on multi-head attention neural networks and includes convolutional networks, allowing high prediction performance and the ability to estimate confounders. The developed framework presents an innovative methodology for validating potential causes: through a second neural network model, a direct and inverse model is trained with the potential causal lag between the causal variable under study and the target variable. In addition, another potential causal discovery method based on a non-linear correlation test is implemented. The main results show that the proposed model outperforms other causal algorithms concerning the F1 Score and achieves a considerable performance in finding temporal lags. In addition, the network fit (evaluated through the mean squared error) is significantly improved when only the causal variables predicted by the proposed algorithm (with their corresponding causal lags) are used for model training, relative to the incorporation of the entire time series dataset for the prediction of the target variable.