Causal AI database: first graph, then effect (RnD heading)
Causal AI database: first graph, then effect (Rubric #RnD)
Watched a short report by Vadim Porvatov from SberChallenges and prospects of Causal AIfrom the Data Fest 2026. It’s a good fifteen-minute route on a topic that’s often boiled down to “correlation doesn’t mean causation.” The main thesis here is more practical: before assessing the effect of an action, it is necessary to restore at least a plausible structure of causes.
Vadim starts with three levels of analytics Descriptive answers the question of what happened. Predictive gives an answer about what will happen. Causal tells you what to change to get the desired result.
The gold standard for the latter is a randomized experiment or a/b test. By the way, I recommend the A/B test.Confidential a/b testing (Trustworthy Online Controlled Experiments)"which I'm talking about." told. But such tests can be expensive, long, unethical, or physically impossible. Then you have to work with the observed data and explicitly record the assumptions.
The report runs through almost the entire map of causal discovery:
From Fisher with randomized experiments to Rubin’s potential outcomes and Judea Pearl’s structural causal models Two stages of causal inference: first identification - which causal graph is acceptable, then estimation - what is the effect of the intervention; Two classic structure search families: constraint-based and score-based; data table input plus expert constraints on the presence and direction of links; Assumptions about acyclicity, Markov property and faithfulness, and most importantly - causal sufficiency: whether all common causes are measured. It is the hidden confounder that can turn a strong correlation into a false story about the effect. PC/SGS algorithms and the FCI family. FCI not only orients the edges, but marks places where the connection can be explained by a hidden variable; FCI-stable, RFCI and FCI+ improve stability, speed or work with sparse graphs. Newer methods are already trying to restore the structure of latent variables.
Separate cold shower - LLM. According to Vadim’s work, models without special configuration do not solve causal discovery poorly, and fine-tuning gives a fragile gain: a small data shift breaks the result. Scenario modeling requires external causal grounding rather than a confident text about “what affects what.”
In the final, Vadim shows CAIMAN, a library being developed in Sber with optimized PC/FCI variants, bootstrap uncertainty assessment and future transfer of computing to the GPU. According to the author, the code is planned to be published; I did not find a public repository at the time of verification.
P.S.
In fact, I was immersed in this topic once when I studied the modern DORA model – its annual research also begins with surveys, that is, with observational, not experimental data. In early methodology, DORA used theoretical hypotheses, latent constructs, factor analysis and PLS-SEM. (That's what D says.ocument 2021 year)
- The report. 2024 year formulated Even closer to the report, researchers ask the DAGs, decide which factors to consider, and then Bayesian estimates of the direction, magnitude, and uncertainty of the effects.
But statistics don't turn respondents' responses into causality at the click. The conclusion remains conditional on the selected graph, the quality of the questions and the confounders taken into account. And here we see a direct link to the report: the most important part of causal inference happens before the coefficients are calculated -- when we decide what variables exist and why there should be an arrow in between.
#AI #Data #MachineLearning #CausalInference #Research #DORA