Toward Causal Scientific Discovery with AI

RIKEN Mathematics & AI Symposium

Transcript

1. RIKEN Mathematics & AI Symposium Toward Causal Scientific Discovery with AI Shohei Shimizu Univ. Osaka, Shiga Univ & RIKEN
2. 2 AI for Science is gaining momentum ◼AI supports the research cycle ◼Data are essential throughout this process ◼Causal discovery connects data to decisions AI-assisted hypothesis generation Experimental and survey design Publishing papers Interpretation and discussion Simulations and Experiments https://www.mext.go.jp/content/20250826-ope_dev02000044427_8.pdf
3. What is Causal Discovery? ◼Assumptions + prior knowledge + data → causal graph → Intervention effects ◼Use data to support causal understanding and intervention decisions Assumptions & Prior knowledge • Hidden common causes • Acyclic or cyclic • Distributions • Functional forms Causal graph Data Inference 3
4. 4 Prediction is not enough ◼Prediction: What will happen? ◼Causality: What would happen if we intervened? ◼Causal understanding supports better decisions Weight Weight … . . . ෣ Disease risk LDL … ෣ Disease risk LDL . . . Blood pressure Blood pressure Prediction Causality
5. LLM supplies prior knowledge ◼Causal parrots (Zečević+2023) ◼LLM + statistical causal discovery (Takayama+2025) No prior knowledge Performance degrades with limited data With LLM-based prior knowledge Health checkup data unseen by the LLM 5
6. 6 Methodology
7. 7 Framework of Causal Discovery ◼Structural causal model (Pearl, 2009) 𝑥𝑖 = 𝑓𝑖 (parents of 𝑥𝑖 , 𝑒𝑖 ) 𝑥3 𝑒3 𝑥1 𝑒1 𝑥2 𝑒2 Error terms ◼Intervention effect 𝑝(𝑥2 |𝑑𝑜 𝑥1 = 𝑐 ) := 𝑝(𝑥𝑖 |Replace the equation for 𝑥1 with 𝑥1 = 𝑐) 𝑥3 𝑥1 𝑐 𝑥2 𝑒2 ◼Make assumptions about the causal model and search for models consistent with the observed data
8. Basic Idea of Causal Discovery (Spirtes et al., 1993; 2001) 1. Assumptions: Acyclicity + No hidden common causes 2. Find the graph that best matches the data among such causal graphs that satisfy the assumptions Three candidates (a) 𝑥1 𝑥2 (b) 𝑥1 𝑥2 (c) 𝑥1 𝑥2 If 𝑥1 and 𝑥2 are independent in the data, select (c) If 𝑥1 and 𝑥2 are dependent in the data, select (a) and (b), but (a) and (b) are indistinguishable 8
9. 9 Identifiability matters ◼Both models (a) and (b) represent the same dependency structure despite having different causal structures Two candidates if 𝑥1 and 𝑥2 are dependent (a) 𝑥1 𝑥2 (b) 𝑥1 𝑥2 ◼Cannot conclude that, for example, (a) is correct without considering identifiability
10. Additional information can break equivalence ◼Under linearity and non-Gaussianity, two models identifiable (Shimizu+2006) Two candidates if 𝑥1 and 𝑥2 are dependent (a) 𝑥1 𝑥2 𝑥1 (b) 𝑥2 ◼Wrong causal direction makes explanatory variable and regression residual dependent 𝑥2 = 𝑏21 𝑥1 + 𝑒2 = 𝑏21 𝑒1 + 𝑒2 (2) 𝑟1 = 𝑥1 − = 1− cov(𝑥1 ,𝑥2 ) 𝑥2 var(𝑥2 ) 𝑏21cov(𝑥1 ,𝑥2 ) var(𝑥2) 𝑏 cov(𝑥 ,𝑥 ) 𝑒1 − 21var(𝑥 1) 2 𝑒2 2 10
11. 11 Beyond Basic Linear Model ◼Additive nonlinearity with hidden common causes (Maeda+2023; Pham+2026) 𝑥𝑖 = ෍ 𝑓𝑖𝑗 (𝑥𝑗 ) + 𝑗: 𝑝𝑎𝑟(𝑥𝑖 ) ෍ 𝑔𝑖𝑘 (𝑢𝑘 ) + 𝑒𝑖 𝑘: 𝑝𝑎𝑟(𝑥𝑖 ) 𝑥4 𝑥4 𝑥1 𝑥3 𝑥1 𝑥2 𝑥3 𝑥2 ◼Beyond additivity: Location-scale noise (Khan+2026) 𝑥𝑖 = ෍ 𝑗: 𝑝𝑎𝑟(𝑥𝑖 ) 𝑓𝑖𝑗 (𝑥𝑗 ) + ෍ 𝑗: 𝑝𝑎𝑟(𝑥𝑖 ) ℎ𝑖𝑗 (𝑥𝑗 ) ෍ 𝑘: 𝑝𝑎𝑟(𝑥𝑖 ) 𝑔𝑖𝑘 (𝑢𝑘 ) + ෍ 𝑗: 𝑝𝑎𝑟(𝑥𝑖 ) ℎ𝑖𝑗 (𝑥𝑗 ) ෍ 𝑘: 𝑝𝑎𝑟(𝑥𝑖 ) 𝑚𝑖𝑘 (𝑢𝑘 ) 𝑒𝑖
12. 12 Causal Direction Identification with Hidden Common Causes ◼Three models identifiable (Hoyer+2008) 𝜆21 𝑒2 𝑥2 𝑢1 𝜆21 𝜆11 𝑏21 𝑥1 1 𝑥2 = 𝑏21 𝑥1 𝑒1 𝑒 0 𝜆11 𝑒1 2 1 𝜆21 𝑢 1 𝑒2 𝑥2 𝑢1 𝜆21 𝜆11 𝑏12 𝑥1 𝑥1 1 𝑏12 𝑥2 = 0 1 𝑒1 𝑒2 𝑒 𝜆11 𝑒1 2 𝜆21 𝑢 1 𝑥2 𝑢1 𝜆11 𝑥1 𝑒1 𝑒 𝑥1 1 0 𝜆11 𝑒1 𝑥2 = 0 1 𝜆21 2 𝑢1 ◼Causal ordering identifiable in multivariate cases (Salehkaleybar+2020) ◼Partially identifiable in nonlinear models using relationships with other variables (Pham+2026)
13. 13 Applications
14. Application Example: Preventive Medicine (Okuda+2026) ◼Long-term effects of health guidance on health outcomes ◼Longitudinal health records and workflow constraints Health guidance Estimated causal graph Health outcomes Bootstrap distribution of intervention effects Lifestyle & medication Gender & Age 14
15. 15 Application Example: Materials Science ◼Why do variables help prediction? ◼Causal or correlational? Descriptor Material property or Descriptor Material property or Common causes Descriptor Material property ?
16. Application Example: Root Cause Analysis Beyond Anomaly Detection ◼Machine learning detects anomalies ◼Causal discovery identifies root causes (Fujiwara+2025) Root cause Anomaly indicator 16
17. 17 Beyond Given Variables Choosing the variables can itself be part of the problem
18. Conventional causal discovery assumes that the variables are given ◼Variables are not always given ◼Too fine: sensor-level ◼Too coarse: total cholesterol ◼Need to discover variables + causal relations kneader cutter chiller Chalupka et al. (2017) 18
19. Causal representation learning (Scholkopf et al., 2021) ◼Learn latent variables from observed data and discover causal relations among them 𝒙=𝑓 𝒔 𝑠𝑖 = 𝑓𝑖 (parents of 𝑠𝑖 , 𝑒𝑖 ) ◼One approach to relaxing the given-variable assumption 𝑠1 𝑠3 𝑠2 Adapted from Shimizu+2009 19
20. Group-Level Causal Discovery ◼Linear Non-Gaussian cases (Kawahara et al., 2010) • Assume that the causal relations among groups form a DAG • Estimate both the grouping and the causal graph among the groups • Variables that share hidden common causes can be grouped (b) Variable-Level Causal Graph (b) Group-Level Causal Graph 20
21. 21 What Is Still Missing? ◼Causal variables may be observed, latent, or constructed from groups of variables ◼Groups and learned latent variables may lack clear scientific meaning ◼Their scientific meaning is a separate question Chalupka et al. (2017)
22. Open problem: What Makes a Variable Scientifically Meaningful? ◼Possible criteria • Appropriate granularity (level of detail) • Can we intervene on it? ◼Criteria depend on scientific judgment and the research question 22
23. 23 Causal Variable and Structure Discovery ◼Discover causal variables at multiple levels of granularity • Find and organize valid groupings hierarchically • Also consider hierarchies of latent factors ◼Provide data-driven candidates for scientific judgment Kawahara+2010 Zeng+2021
24. 24 Summary and Future Direction ◼Summary • Prediction → Causal understanding • Given variables → Causal variable and structure discovery ◼Future direction • Bring these ideas into the AI for Science cycle • Data-driven study of causal relations and intervention effects → New hypotheses, new variables, new experiments