Decoupling Intrinsic and Measurement Trends: A Crucial Consideration in Time Series Causal Discovery
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 trend, measurement error, data preprocessing
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: In this paper, we introduce a novel algorithm TrendDiff that effectively distinguishes between intrinsic (real) trends and measurement (false) trends, leading to enhanced detrending and improved causal discovery.
Abstract: In the realm of time series data, it is common to encounter time trends, which manifest as a function concerning time within a given data span. Time trends can be classified into intrinsic (real) and measurement (false) trends. Intrinsic trends are inherent to the underlying mechanisms of the variables, while measurement trends are essentially measurement errors unique to the observed values (e.g., an increase in diagnosed thyroid nodule patients due to enhanced medical techniques, despite a stable incidence rate over time). Measurement trends can critically influence the results of a variety of causal discovery methods and hence, necessitate elimination prior to causal analytic procedures. In this study, we introduce a novel framework capable of detecting all trend-influenced variables and distinguishing between intrinsic and measurement trends, called Trend Differentiator (TrendDiff). This approach consists of two primary steps: trend variable identification and trend type differentiation. The first step leverages Constraint-based Causal Discovery from heterogeneous/Nonstationary Data (CD-NOD) to identify variables with trends. Following this, we utilize the structure characteristics to differentiate between intrinsic and measurement trends. Experimental results on various synthetic scenarios and real-world data sets are employed to demonstrate the efficacy of our methods.
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: 7050
Loading