Learning Time-shared Hidden Heterogeneity for Counterfactual Outcome Forecast

Jie Peng, Hao Zou, Renzhe Xu, Haotian Wang, Peng Cui

Keywords: Hidden Heterogeneity; Counterfactual Outcome Forecast; Time series

Abstract: Forecasting counterfactual outcome in the longitudinal setting can be critical for many time-related applications. To solve this problem, the previous works propose to apply different sequence models including long short-term memory (LSTM) networks and transformers to model the relationship between the observed histories, treatments and outcomes, and apply various approaches to remove treatment selection bias. However, these methods neglect the hidden heterogeneity of outcome generation among samples induced by hidden factors which can bring hurdles to counterfactual outcome forecast. To alleviate this problem, we capture the hidden heterogeneity by recovering the hidden factors and incorporate it into the outcome prediction process. Specifically, we propose a Time-shared Heterogeneity Learning from Time Series (THLTS) method which infers the shared part of hidden factors characterizing the heterogeneity across time steps with the architecture of variational encoders (VAE). This method can be a flexible component and combined with arbitrary counterfactual outcome forecast method. Experimental results on (semi-)synthetic datasets demonstrate that combined with our method, the mainstream models can improve their performance.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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: 12869