Causal Discovery on Real Data: Combining Causal Graphs, Uplift and LLMs

Dmitry Kornilov23 May 2026

Problem

A bank targets communications (SMS, push offers) at the clients who should gain most. Uplift models rank clients by expected benefit, but they are trained on whatever features are available. A control feature that is a collider or a mediator biases the effect estimate, and a missing common cause leaves a spurious effect. The business then sees implausible results and stops trusting the model.

Example from the article: a model trained on a "historical" feature set estimated the effect of credit card approval on client debt at −0.1748, while the true effect was +1.20, so the sign flipped.

Solution in brief

Discover the causal graph from the data with a constraint-based algorithm (PC, FCI) and orient the edges the algorithm leaves undirected using an LLM or an analyst. Then take the adjustment set from the graph with DoWhy and train the uplift model on exactly those features. The team made graph discovery practical on large feature sets by parallelizing the independence tests in its CAIMAN library.

Questions this case answers
  • How to select adjustment variables for a treatment-effect model from observational data when the candidate feature set is large and some features are colliders or mediators
  • How to recover a causal graph from observational data when the discovery algorithm leaves some edges undirected and no expert is available to orient them
  • How to stop uplift models from being biased by wrong control variables
  • How to choose which features to control for when estimating a treatment effect from observational data

Tools and methods

  • PC algorithm and FCI — constraint-based causal discovery; FCI also flags missing variables
  • CAIMAN — the team's library with parallelized discovery algorithms and the new B-RAI and B-FCI
  • LLM edge orientation — pairwise questions about undirected or implausible edges
  • DoWhy — selects the adjustment set from a NetworkX graph
  • CausalML and CatBoost metalearners — uplift models and ATE estimation
Continue reading — it's free

Google or a one-time email code. No password.