Discovering Granger Causality With Convolutional Neural Networks

Abstract

This study explores identifying unidirectional Granger causality between a single variable and the rest of the variables with Convolutional Neural Networks (CNN) for a time series data. A novel approach is suggested in which CNN kernel weights are used as Granger causality coefficients. The question whether or not the near future occurrence probability of a selected variable can be found using the past occurrences of itself and other variables is answered. Since the proposed method enables the usage of gradient descent with Graphics Processing Unit (GPU) power, it paves the way for calculating Granger causality with more variables. Although some other Deep Learning techniques have been utilized in causality discovery, this kind of CNN usage is a new idea and F1-score of 0.8399 obtained with a real alarm dataset logged by industrial machinery suggests that it is a successful method. While the proposed approach is generic and applicable to any time series data in any area from finance to healthcare or manufacturing industry, for any problem specific target like classification or regression, such an alarm prediction model could be convenient to take operational actions in manufacturing facilities where predictive maintenance with sensor measurements is limited and only alarm logs are available.