causal

**Causal Inference Deep Learning** is **methods for learning causal relationships from data and predicting effects of interventions using neural networks combined with causal modeling frameworks** — moves beyond correlation to causation. Causal understanding essential for science and policy. **Causal Graphs and DAGs** directed acyclic graphs represent causality: edges = causal arrows. Confounders: common causes of two variables. Colliders: common effects. Structure determines valid inference. **Confounding** unobserved confounder affects treatment and outcome, biasing causal estimates. **Causal Discovery** learn graph structure from observational data. PC algorithm (constraint-based), FCI (handles latent confounders), score-based methods. Identifiability challenging without assumptions. **Causal Inference from Observational Data** estimate treatment effect without randomization. **Potential Outcomes Framework** Rubin Causal Model: for each unit, two potential outcomes Y(1) (treated) and Y(0) (untreated). Observed one, other counterfactual. **Average Treatment Effect (ATE)** E[Y(1) - Y(0)] over population. **Propensity Score Matching** estimate probability of receiving treatment given covariates (propensity score). Match treated/untreated with similar scores. Removes confounding from measured covariates. **Doubly Robust Methods** combine regression and propensity score models. Robust if either correct. **Causal Forests** random forests estimating heterogeneous treatment effects: different people respond differently. Conditional Average Treatment Effect (CATE) varies with features. **Deep Learning for Causal Inference** neural networks as flexible function approximators in causal methods. Estimate propensity scores, outcomes, heterogeneous effects. **Instrumental Variables** confounder unobserved. Use instrument Z: affects treatment but only through treatment (exclusion restriction). Allows causal inference. **Causal Representation Learning** learn representations that disentangle causes and effects. **Counterfactual Explanations** for prediction x, what changes make prediction change? Minimally perturbed input with different prediction. **Do-Calculus** Pearl's framework: transform conditional probabilities to interventional probabilities. Rules determine identifiability. **Backdoor Criterion** conditions for causal identification adjusting for confounders. **Frontdoor Criterion** identifies causal effect when backdoor open, frontdoor closed. Requires mediator. **Structural Causal Models (SCM)** directed acyclic graphs + functional relationships + noise. **Latent Confounders** unobserved confounders. Methods: instrumental variables, causal graphs with latent variables. **Time Series Causality** Granger causality: past X predicts Y better than Y alone. Not true causality but useful for sequences. **Mediation Analysis** decompose effect into direct (unmediated) and indirect (through mediator). **Sensitive Analysis** test robustness of causal estimates to unobserved confounding. **Fairness and Causality** bias in predictions due to discriminatory causal relationships. Interventional fairness: outcomes fair under intervention, not just association. **Causal Explanation** predict outcome, explain via causal pathways. Saliency + causality. **Applications** medical treatment effect estimation, economics (policy evaluation), marketing (campaign effectiveness), recommendation systems. **Challenges** identifiability: multiple models consistent with data. Assumptions often untestable. **Software and Tools** PyMC3, Stan for Bayesian causal inference. DoWhy library for causal methods. **Causal Deep Learning combines neural network flexibility with causal frameworks** enabling better science and policy decisions.