[1J4-GS-2] Machine learning: Fundamental theory (1)

[1J4-GS-2-04] Estimation of interventinal effect on prediction models for time series data

〇Keisuke Kiritoshi¹, Wataru Kurebayashi², Tomonori Izumitani¹, Kazuki Koyama¹, Daichi Kimura¹, Tomomi Okawachi¹, Shohei Shimizu² (1. NTT Communications Corporation, 2. Shiga University)

Keywords:causality analysis, machine Learning, Interventional effect

For prediction models of machine learning including operation variables, there are applications to specify operation variables with high contribution from the weights and partial derivatives of models and to connect them to actual operations. On the other hand, since prediction models do not consider the causal relationship behind the data, the intervention effect on the model may differ from the actual system even if the prediction accuracy is high. Blobaum et al. propose the framework which obtains the intervention value which approaches the prediction value of the prediction model to the desired value through the causal graph by combining the regression model and causal graph. In this paper, we propose a method to obtain the intervention effect for the time series model and the time series causal graph using this framework. We evaluate the error between the predicted value and the desired value by the intervention obtained only by the regression model and the proposed method for the time series data by experiments of artificial data and simulation data.