Causal discovery – real-world applications, challenges, and open problems

Before we wrap up this chapter, let’s take a broader perspective and discuss the applicability of causal discovery to real-world problems and challenges that may arise along the way.

In the previous chapter, we mentioned that Alexander Reisach and colleagues have demonstrated that the synthetic data used to evaluate causal discovery methods might contain unintended regularities that can be relatively easily exploited by these models (Reisach et al., 2021). The problem is that these regularities might not be present in real-world data.

Another challenge is that real-world data with a known causal structure is scarce. This makes synthetic datasets a natural benchmarking choice, yet this choice leaves us without a clear understanding of what to expect of causal structure learning algorithms when applied to real-world datasets.

The lack of reliable benchmarks is one of the main challenges in...

Wrapping it up!

In this chapter, we introduced several methods and ideas that aim to overcome the limitations of traditional causal discovery frameworks. We discussed DECI, an advanced deep learning causal discovery framework, and demonstrated how it can be implemented using Causica, Microsoft’s open source library, and PyTorch.

We explored the FCI algorithm, which can be used to handle data with hidden confounding, and introduced other algorithms that can be used in similar scenarios. These methods provide a strong foundation for tackling complex causal inference problems.

After that, we discussed two frameworks, ENCO and ABCI, that allow us to combine observational and interventional data. These frameworks extend our ability to perform causal discovery and provide valuable tools for data analysis.

Finally, we discussed a number of challenges that we face when applying causal discovery methods to real-world problems.

We are inexorably approaching the end of our...

References

Andrews, B., Sprites, P., & Cooper, G. F. (2020). On the Completeness of Causal Discovery in the Presence of Latent Confounding with Tiered Background Knowledge. International Conference on Artificial Intelligence and Statistics.

Cai, R., Qiao, J., Zhang, K., Zhang, Z., & Hao, Z. (2021). Causal discovery with cascade nonlinear additive noise models. ACM Trans. Intell. Syst. Technol., 6(12).

Geffner, T., Antorán, J., Foster, A., Gong, W., Ma, C., Kıcıman, E., Sharma, A., Lamb, A., Kukla, M., Pawlowski, N., Allamanis, M., & Zhang, C. (2022). Deep End-to-end Causal Inference. arXiv.

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2020). Generative adversarial networks. Communications of the ACM, 63(11), 139-144.

Goudet, O., Kalainathan, D., Caillou, P., Guyon, I., Lopez-Paz, D., & Sebag, M. (2018). Causal generative neural networks. arXiv.

Huang, Y., Kleindessner...

Extra – going beyond observations

In certain cases, we might be able to intervene on some or all variables in order to facilitate or improve the results of a causal discovery process.

In this short section, we’ll introduce two methods that can help us make sure that we make good use of such interventions.

ENCO

Efficient Neural Causal Discovery (ENCO; Lippe et al., 2022) is a causal discovery method for observational and interventional data. It uses continuous optimization and – as we mentioned earlier in the section on DECI – parametrizes edge existence and its orientation separately. ENCO is guaranteed to converge to a correct DAG if interventions on all variables are available, but it also performs reasonably well on partial intervention sets. Moreover, the model works with discrete, continuous, and mixed variables and can be extended to work with hidden confounding. The model code is available on GitHub (https://bit.ly/EncoGitHub).

ABCI

...

Causal discovery – real-world applications, challenges, and open problems

Before we wrap up this chapter, let’s take a broader perspective and discuss the applicability of causal discovery to real-world problems and challenges that may arise along the way.

In the previous chapter, we mentioned that Alexander Reisach and colleagues have demonstrated that the synthetic data used to evaluate causal discovery methods might contain unintended regularities that can be relatively easily exploited by these models (Reisach et al., 2021). The problem is that these regularities might not be present in real-world data.

Another challenge is that real-world data with a known causal structure is scarce. This makes synthetic datasets a natural benchmarking choice, yet this choice leaves us without a clear understanding of what to expect of causal structure learning algorithms when applied to real-world datasets.

The lack of reliable benchmarks is one of the main challenges in...

Wrapping it up!

In this chapter, we introduced several methods and ideas that aim to overcome the limitations of traditional causal discovery frameworks. We discussed DECI, an advanced deep learning causal discovery framework, and demonstrated how it can be implemented using Causica, Microsoft’s open source library, and PyTorch.

We explored the FCI algorithm, which can be used to handle data with hidden confounding, and introduced other algorithms that can be used in similar scenarios. These methods provide a strong foundation for tackling complex causal inference problems.

After that, we discussed two frameworks, ENCO and ABCI, that allow us to combine observational and interventional data. These frameworks extend our ability to perform causal discovery and provide valuable tools for data analysis.

Finally, we discussed a number of challenges that we face when applying causal discovery methods to real-world problems.

We are inexorably approaching the end of our...

References

Andrews, B., Sprites, P., & Cooper, G. F. (2020). On the Completeness of Causal Discovery in the Presence of Latent Confounding with Tiered Background Knowledge. International Conference on Artificial Intelligence and Statistics.

Cai, R., Qiao, J., Zhang, K., Zhang, Z., & Hao, Z. (2021). Causal discovery with cascade nonlinear additive noise models. ACM Trans. Intell. Syst. Technol., 6(12).

Geffner, T., Antorán, J., Foster, A., Gong, W., Ma, C., Kıcıman, E., Sharma, A., Lamb, A., Kukla, M., Pawlowski, N., Allamanis, M., & Zhang, C. (2022). Deep End-to-end Causal Inference. arXiv.

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2020). Generative adversarial networks. Communications of the ACM, 63(11), 139-144.

Goudet, O., Kalainathan, D., Caillou, P., Guyon, I., Lopez-Paz, D., & Sebag, M. (2018). Causal generative neural networks. arXiv.

Huang, Y., Kleindessner...