Dynamic causality

March 5, 2024 — March 5, 2024

graphical models
how do science
machine learning
neural nets
Figure 1

I am not yet sure if this is a real thing. I came across some interesting terminology that argues that sometimes we might want to treat streams of things as being generated by a small evolving graphical model rather than by a large static one.

That is certainly my reading of DeepCausality, which seems to target streaming recommender-system-like problems. However, that page has some red flags, such as a lack of equations, opaque terminological name-drops, lack of citations and non-explanations:

DeepCausality uses the causaloid as its central structure, an idea borrowed from a novel causal concept pioneered by Lucien Hardy at the Perimeter Institute of Theoretical Physics. The causaloid encodes a causal relation as a causal function that maps input data to an output decision. It then determines whether, on the input data, the causal relation, encoded as a function, holds.

That is a provocative start. It is also all there is. Here are the articles on causaloids which I assume they refer to: Hardy (2005), Hardy (2007). Connecting these articles on quantum gravity to inference in recommender systems is left as an exercise.

1 References

Hardy. 2005. Probability Theories with Dynamic Causal Structure: A New Framework for Quantum Gravity.”
———. 2007. Towards Quantum Gravity: A Framework for Probabilistic Theories with Non-Fixed Causal Structure.” Journal of Physics A: Mathematical and Theoretical.
Rubenstein, Bongers, Schölkopf, et al. 2018. From Deterministic ODEs to Dynamic Structural Causal Models.” In Uncertainty in Artificial Intelligence.
Xu, Tan, Fu, et al. 2022. Dynamic Causal Collaborative Filtering.” In Proceedings of the 31st ACM International Conference on Information & Knowledge Management. CIKM ’22.