Abstract
Over the past years, researchers have proposed various methods to discover causal relationships among time-series data as well as algorithms to fill in missing entries in time-series data. Little to no work has been done in combining the two strategies for the purpose of learning causal relationships using unevenly sampled multivariate time-series data. In this paper, we examine how the causal parameters learnt from unevenly sampled data (with missing entries) deviates from the parameters learnt using the evenly sampled data (without missing entries). However, to obtain the causal relationship from a given time-series requires evenly sampled data, which suggests filling the missing data values before obtaining the causal parameters. Therefore, the proposed method is based on applying a Gaussian Process Regression (GPR) model for missing data recovery, followed by several pairwise Granger causality equations in Vector Autoregressive form to fit the recovered data and obtain the causal parameters. Experimental results show that the causal parameters generated by using GPR data filling offers much lower RMSE than the dummy model (fill with last seen entry) under all missing values percentage, suggesting that GPR data filling can better preserve the causal relationships when compared with dummy data filling, thus should be considered when dealing with unevenly sampled time-series causality learning.
Introduction
In industrial and real-world settings, sensors fail, networks lag, and data is often missing. Causal discovery—the process of determining if one variable influences another over time—is a cornerstone of predictive maintenance and systems engineering. The standard tool for this is Granger Causality, which essentially asks: "Does knowing the past of variable $A$ help me predict the future of variable $B$ better than just knowing the past of $B$ alone?"
However, Granger Causality and the Vector Autoregressive (VAR) models it relies on require evenly sampled data. In the real world, data is rarely perfect. Sensors might only report when a value changes, or hardware limitations might lead to frequent dropouts. Traditionally, researchers either ignored the missing data or used "dummy" filling (carrying the last known value forward). This paper investigates how these gaps distort the "causal parameters" of a system and proposes a more robust way to patch the holes using Gaussian Process Regression (GPR).
Methodology
The researchers developed a three-stage pipeline to test their hypothesis:
- Data Thinning: They took a complete, evenly sampled dataset and randomly dropped between 10% and 80% of the entries to simulate real-world "messy" data.
- Imputation (Data Filling): They compared two methods to recover the missing values: Gaussian Process Regression (GPR), a non-parametric, probabilistic approach using a Radial Basis Function (RBF) kernel, and Dummy Filling, a baseline method where a missing entry is replaced by the last observed value.
- Causal Analysis: The recovered datasets were fed into a VAR model to calculate pairwise Granger Causality parameters. These were compared against the "ground truth" parameters calculated from the original data using Root Mean Square Error (RMSE).
The study utilized the PHM08 (Prognostics and Health Management) dataset from NASA, specifically simulated data from a turbofan engine degradation model (C-MAPSS). 15 sensor data streams from the first 11 engines were selected for analysis.
Results and Discussion
The quantitative results were decisive. Across all missing data percentages (from 10% to 80%), GPR-filled data resulted in causal parameters with a significantly lower RMSE compared to dummy filling.
As shown in the figure above, the error for the dummy model spikes aggressively as the percentage of missing data increases. In contrast, the GPR model's error grows much more slowly, showing remarkable resilience even when 80% of the data points are gone. Visualizations further support these findings, showing that GPR predictions closely track the ground truth curves, whereas dummy filling often misses the peaks and troughs entirely, leading to "flat" segments that mask the system's dynamics.
The paper highlights the "Staircase" problem: dummy filling creates artificial plateaus that destroy the subtle temporal correlations needed for Granger Causality. GPR, by contrast, provides a smoothing effect that helps reduce noise and follows the "true" trend of the time-series more closely than the jagged edges of raw, thinned data.
Conclusion
Causal parameters are highly sensitive to how data is handled. Even if a filling method looks acceptable visually, it can lead to completely different conclusions about which variables are driving the system. While GPR is computationally expensive ($O(N^3)$ complexity), it significantly outperforms simple imputation methods in preserving causal signals. Future work could explore optimized "Sparse GPs" for larger datasets or the use of different kernels to better capture specific time-series characteristics.
Summary updated