Causal AI

Background

  • There is an emerging view that that current efforts towards interpretability and robustness are fruitless and only an incorporation of causality can provide answers (Pearl and Mackenzie 2018).
  • Pearl (2019) argues that AI is current stuck at the association level: models are limited to learning \(P(y|X)\) (“glorified curve fitting”). Starting from causal graphical models (CGM) improves transparency and domain adaptability.

Structure learning

  • Zheng et al. (2018) proposes to cast the combinatorial problem of learning a CGM into a continuous problem that can be learned through standard non-convex constrained optimization for linear structural equation models (SEM).
  • Lachapelle et al. (2019) extend this idea to the non-linear case.
  • Bussmann, Nys, and Latré (2020) propose Neural Additive Vector Autoregression (NAVAR) for (Granger) causal discovery in time series setting. The model can be seen as a Generalised Additive Model and is therefore inherently (somewhat) interpretable. It is based on the assumption that contemporary dependencies between variables are linear, only dependencies through time require non-linear model.

Link to CE and algorithmic recourse

  • Joshi et al. (2019) make an interesting link between CGM and counterfactual explanations: they draw an analogy between hidden confounders in CGMs and the latent manifold which REVISE traverses to propose recourse. Run a single experiment on TWIN dataset and show that recommended recourse changes qualitatively as confounding is introduced.
  • Karimi et al. (2020) develop two probabilistic approaches to algorithmic recourse in the case of limited causal knowledge.
    • In essence the probabilistic approach boils down to assuming a Gaussian Process prior for the causal mapping parent nodes to node \(X\). This yields a posterior noise distribution, which in turn can be used to draw from a counterfactual distribution.
  • Karimi, Schölkopf, and Valera (2021) demonstrate how to go from counterfactuals to interventions in the case of complete knowledge of the CGM. Propose a shift of paradigm from recourse via counterfactuals to recourse through minimal interventions.
  • Can explore the link between CGM and CE further, perhaps in the context of Bayesian classifier (Schut et al. 2021).
    • In particular, it might be possible to draw an analogue between Schut et al. (2021) (low epistemic + aleatoric uncertainty) and the counterfactual distribution proposed by Karimi et al. (2020). Could further try to account for hidden confounders as in Joshi et al. (2019).
  • Karimi, Schölkopf, and Valera (2021) can be solved through by building on existing frameworks for generating nearest counterfactual explanations - could try to apply Schut et al. (2021)?
  • Applications at ING?
    • Apply to loan application decision system (if exists)
    • Apply to credit scoring (perhaps even Dutch government scandal)
    • …

References

Bussmann, Bart, Jannes Nys, and Steven Latré. 2020. “Neural Additive Vector Autoregression Models for Causal Discovery in Time Series Data.” arXiv Preprint arXiv:2010.09429.
Joshi, Shalmali, Oluwasanmi Koyejo, Warut Vijitbenjaronk, Been Kim, and Joydeep Ghosh. 2019. “Towards Realistic Individual Recourse and Actionable Explanations in Black-Box Decision Making Systems.” arXiv Preprint arXiv:1907.09615.
Karimi, Amir-Hossein, Bernhard Schölkopf, and Isabel Valera. 2021. “Algorithmic Recourse: From Counterfactual Explanations to Interventions.” In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 353–62.
Karimi, Amir-Hossein, Julius Von Kügelgen, Bernhard Schölkopf, and Isabel Valera. 2020. “Algorithmic Recourse Under Imperfect Causal Knowledge: A Probabilistic Approach.” arXiv Preprint arXiv:2006.06831.
Lachapelle, Sébastien, Philippe Brouillard, Tristan Deleu, and Simon Lacoste-Julien. 2019. “Gradient-Based Neural Dag Learning.” arXiv Preprint arXiv:1906.02226.
Pearl, Judea. 2019. “The Seven Tools of Causal Inference, with Reflections on Machine Learning.” Communications of the ACM 62 (3): 54–60.
Pearl, Judea, and Dana Mackenzie. 2018. The Book of Why: The New Science of Cause and Effect. Basic books.
Schut, Lisa, Oscar Key, Rory Mc Grath, Luca Costabello, Bogdan Sacaleanu, Yarin Gal, et al. 2021. “Generating Interpretable Counterfactual Explanations by Implicit Minimisation of Epistemic and Aleatoric Uncertainties.” In International Conference on Artificial Intelligence and Statistics, 1756–64. PMLR.
Zheng, Xun, Bryon Aragam, Pradeep Ravikumar, and Eric P Xing. 2018. “Dags with No Tears: Continuous Optimization for Structure Learning.” arXiv Preprint arXiv:1803.01422.