Understanding cause and effect is crucial in many areas of science and technology, especially when trying to build smart systems that can make decisions or predictions. One of the big challenges researchers face is identifying cause-and-effect relationships when there are hidden factors that are not easily seen in the data. Traditional methods often rely on just observational data (things we can see and measure), but this approach can miss important details when the system is complex.

In our research, we introduce a new method called CAnDOIT — CAusal Discovery with Observational and Interventional data from Time series — that can better uncover cause-and-effect relationships by using both observational and interventional data: what we observe and what happens when we deliberately intervene in the system. This method is particularly useful in real-world applications like robotics, where simply observing things is often not enough to figure out what causes what.

We tested this method using both randomly generated models and a well-known benchmark for causal structure learning in a robotic manipulation environment. The results show that CAnDOIT can improve the accuracy of causal analysis by incorporating both types of data.

To make this method accessible, we have made a Python implementation of it publicly available on GitHub: https://github.com/lcastri/causalflow.

PDF: https://onlinelibrary.wiley.com/doi/epdf/10.1002/aisy.202400181
Full Article: https://onlinelibrary.wiley.com/doi/full/10.1002/aisy.202400181

BibTeX
@article{castri2024candoit,
title={CAnDOIT: Causal Discovery with Observational and Interventional Data from Time Series},
author={Castri, Luca and Mghames, Sariah and Hanheide, Marc and Bellotto, Nicola},
journal={Advanced Intelligent Systems},
pages={2400181},
publisher={Wiley Online Library}
}