Using causal discovery for feature selection in multivariate numerical time series

Abstract

Time series data contains temporal ordering, which makes its feature selection different from the normal feature selection. Feature selection in multivariate time series has two tasks: identifying the relevant features and finding their effective window sizes of lagged values. The methods extended from normal feature selection methods do not solve this two-dimensional feature selection problem since they do not take lagged observations of features into consideration. In this paper, we present a method using the Granger causality discovery to identify causal features with effective sliding window sizes in multivariate numerical time series. The proposed method considers the influence of lagged observations of features on the target time series. We compare our proposed feature selection method with several normal feature selection methods on multivariate time series data using three well-known modeling methods. Our method outperforms other methods for predicting future values of target time series. In a real world case study on water quality monitoring data, we show that the features selected by our method contain four out of five features used by domain experts, and prediction performance on our features is better than that on features of domain experts using three modeling methods.