Mátyás Schubert

Local Causal Discovery for Statistically Efficient Causal Inference Causal discovery methods can identify valid adjustment sets for causal effect estimation for a pair of target variables, even when the underlying causal graph is unknown. Global causal discovery metho...

LOAD is already my second work with the team of Tom Claassen and @smaglia.bsky.social 🥳 Check out the details of the paper at arxiv.org/abs/2510.14582 and load optimal adjustment sets without waiting using the publicly available code at github.com/Matyasch/load!

On both synthetic and realistic data LOAD
🏎️is more computationally efficient than global methods, performing close to local methods,
💎recovers high-quality, statistically efficient adjustment sets,
🔮thus enables reliable causal effect estimation even at scale

7/8

LOAD follows 5 steps:
➡Learn causal relations between targets
✅Test identifiability of the effect
🐣Find explicit descendants of treatment
🧩Find mediators
🎯Collect optimal adjustment set
For unidentifiable effects, LOAD exits early and returns locally valid adjustments

6/8

To do this, we develop a sufficient and necessary test for the identifiability of the causal effect of a treatment on an outcome using only local information around the treatment and its siblings, no matter how far the treatment and the outcome are in the causal graph 🔭

5/8

🎯 Local Optimal Adjustments Discovery (LOAD) does exactly that! It provably finds the same ✨optimal adjustments✨ as global methods, but using much more ⚡computationally efficient⚡ local causal discovery around variables

4/8

🌐 Global discovery methods can find optimal adjustment sets, but at a huge computational cost.
📍 Local discovery methods are fast, but can only find sub-optimal adjustment sets.

Can we get the best of both worlds and find optimal adjustment sets from local information?

3/8

While all valid adjustment sets enable unbiased estimation of causal effects, using the optimal adjustment set in terms of asymptotic variance is crucial for reliable causal effect estimation!⚠️

But how to find the optimal adjustment set if the causal graph is not available?

2/8

Estimating causal effects efficiently doesn’t have to mean discovering the entire causal graph! Now you can find the optimal adjustment from only local information using LOAD!

📜 Preprint: arxiv.org/abs/2510.14582
👾 Code: github.com/Matyasch/load
🧵 1/8

Carlos Cinelli, Avi Feller, Guido Imbens, Edward Kennedy, Sara Magliacane, Jose Zubizarreta
Challenges in Statistics: A Dozen Challenges in Causality and Causal Inference
https://arxiv.org/abs/2508.17099

Are you interested in improving the #interpretability, #robustness and #safety of current AI systems with #causality and #RL?

Apply to our PhD position in Amsterdam 🚲🌷🇳🇱

Deadline: June 15

CAR Causal Abstractions and RepresentationsWorkshop @ UAI 2025July 25th 2025, Rio de Janeiro 🇧🇷

🔥 Got a great work on causal representation learning, abstraction, high-dimensional discovery, or other hot topics in causality?

🇧🇷 Don’t miss your chance to present in Rio at the CAR Workshop at #UAI2025!

⏰ Deadline is in 1 week – May 26!
🌐 sites.google.com/view/car-25/

A little over a week ago, I had the chance to attend #AISTATS and present our poster on SNAP (matyasch.github.io/snap)! Three days of brilliant invited talks and a stream of fascinating papers left me with a much longer reading list about ideas to explore.

Vacancy — PhD Position on Learning Concepts with Theoretical Guarantees Using Causality and RL Are you interested in improving the interpretability, robustness and safety of current AI systems? If the answer is yes, please continue reading!

New PhD position at the University of Amsterdam in @amlab.bsky.social on learning concepts with theoretical guarantees using #causality and #RL with me, Frans Oliehoek (TU Delft) and Herke van Hoof 💥

Deadline: 15 June

werkenbij.uva.nl/en/vacancies...

Just arrived in Phuket for #AISTATS2025. Can't wait to present our poster (in tube) about SNAP 🫰 on day 2, Sunday! Come check it out and let's chat about scalable causal discovery!

⏰ Don't miss Mátyás talk today at 3PM!

🎥 See you online meet.google.com/cqt-ufji-xfz

🤌 ...or live in the CS Department of the University of Pisa!

A few weeks ago, I presented SNAP at the wonderful #Bellairs Workshop on Causality in Barbados🐢

This Friday 🫰meets 🤌 as I will get to present SNAP again at the kick-off of the newest season of @causalclub.di.unipi.it! Check out this, and their other amazing upcoming talks at causalclub.di.unipi.it

Sequential Non-Ancestor Pruning | Matyas Schubert

10/10 SNAP is joint work with a fantastic team of Tom Claassen and @smaglia.bsky.social. Visit our project page on matyasch.github.io/snap/, run SNAP using our publicly available code at github.com/matyasch/snap, and visit to our poster at #aistats2025! 🏖️

9/10 We also evaluate SNAP on semi-synthetic settings including data generated from the MAGIC-NIAB network, which captures genetic effects and phenotypic interactions 🧬 We see that SNAP greatly reduces the number of CI tests and execution time compared to most baselines.

8/10 Many non-ancestors are already identified by marginal tests, enabling prefiltering with SNAP(0) to significantly speed up computation time. Increasing the number of prefiltering iterations k further reduces the number of CI tests needed, especially in dense graphs 🧶

7/10 SNAP(∞) consistently ranks among the best in the number of CI tests and computation time across all domains, while maintaining a comparable intervention distance. In contrast, other methods vary in performance depending on the setting 🚀

6/10 We can also run SNAP until completion, to obtain a stand-alone causal discovery algorithm, called SNAP(∞). SNAP(∞) is sound and complete over the possible ancestors of targets ✅ Thus, unlike previous work on local causal discovery, it finds efficient adjustment sets.

5/10 SNAP is straightforward to combine with readily available causal discovery algorithms 🧩 We can simply stop it at any maximum iteration k and run another algorithm on the remaining variables. We refer to this approach as prefiltering with SNAP(k).

4/10 To solve this task, we show that only possible ancestors of the targets are required to identify their causal relationships and efficient adjustment sets💡 Driven by this, we propose SNAP to progressively prune non-ancestors, leading to much fewer higher order CI tests.

3/10 We formalize this as the task of “targeted causal effect estimation with an unknown graph”, which focuses on identifying causal effects between a small set of target variables in a ✨computationally and statistically efficient way✨

2/10 Discovering causal relations can help us estimate causal effects, but it is expensive 📈 If we are only interested in estimating the causal effects between a few target variables, can we instead only discover a subgraph that includes these targets and their adjustment sets?

Do you want to estimate causal effects for a small set of target variables without knowing the causal graph, but discovering it takes too long? Now you can get adjustment sets in a SNAP🫰accepted at #aistats2025!

📜 arxiv.org/abs/2502.07857
🧩 matyasch.github.io/snap/
🧵 1/10

Professor Imbens also had a mentoring session with our PhD students actively working on causality, discussing their ideas and the potential impact of their applications! 👨‍🔬👩‍🔬

@matyasch.bsky.social @roelhulsman.bsky.social @rmassidda.it @danruxu.bsky.social 🔥

Congrats to @smaglia.bsky.social for now being an ELLIS Scholar! 🤩🥳🎉

if you ever wondered how to concisely represent causal models, don’t miss my talk on causal abstraction tomorrow at @ellisamsterdam.bsky.social