Trend/Seasonality based Causal Structure for Time Series Counterfactual Outcome Prediction

Jung-Hee Kim, Juhyun Lyu, Wonbin Ahn, Sangmin Lee, Sangjun Han, Hyeokjun Choe, Soyeon Park, Woohyung Lim

Primary Area: causal reasoning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: Time series counterfactual outcome prediction, Trend/Seasonality decomposition, Causal structure

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract: In the causal effect estimation, most models have focused on estimating counterfactual outcomes in the static setting, and it is still difficult to predict the outcomes in the longitudinal setting due to time-varying confounder. To resolve the time-varying confounder issue, while the balance representation learning-based approaches have been primarily proposed, they inherently introduces a certain degree of selection bias since the balance representations act as confounders for both treatment and outcomes. In this paper, a new trend/seasonality decomposition based causal structure is proposed for the counterfactual outcome prediction in the time-series setting. We leverage a decomposition methodology to reduce the selection bias further. Specifically, it extracts meaningful decomposed representations such as confounders and adjustment variables, which help to yield more accurate treatment effect estimation with low variance. Inspired by the fact, the proposed causal structure learns trend/seasonality representations as the confounders/adjustment variables in the direction of minimizing the selection bias, and those representations are effective in the counterfactual outcome prediction especially under the long time sequence and high time-varying confounding settings. We evaluate the proposed causal structure with several trend/seasonality decomposition algorithms on synthetic and real-world datasets. From various experiments, the proposed causal structure achieves superior performance over the state-of-the-art algorithms.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4996