Time Series Prediction With Events Disturbance Based Causal Representation Learnin

Xuanzhi Feng

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 prediction; extreme event interference; counterfactual prediction; PUNs networks; causal representation learning

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

Abstract: The value of time series prediction is getting more and more attention, and the prediction of time series data under extreme event disturbance has been difficult, the different distribution of data before and after the event and the different distribution of dataset will lead to the poor prediction accuracy, robustness and generalisation ability of prediction model. In this paper, based on the causal representation learning, we design the SCM structure under event interference and propose the causal representation prediction model, which is divided into two parts, CRP\_Encoder and CRP\_Decoder.CRP\_Encoder completes the extraction of causal representations disturbed by events and those not disturbed by events through the causal factor extractor and the causal representation decoupler; in order to learn the causal mechanism, the equivalence of conditional structure and causal mechanism is proved, and CNN network and causal representation coupler are designed in CRP\_Decoder to learn and predict. The experimental results show that the CRP model has high prediction accuracy, good robustness and strong generalisation ability.

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: 3361