Research map / Probabilistic and scientific modeling / Causal discovery

Causal Dependence Plots

NeurIPS 2024 · Probabilistic and scientific modeling › Causal discovery · approach: Causal effect identification · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes Causal Dependence Plots, which visualize how a black-box model's prediction depends on a predictor while letting other predictors change causally via an auxiliary structural causal model, unifying PDP/ICE with causal effects.

Problem

Standard model-explanation plots (PDP, ICE) hold other predictors fixed or treat them as independent, which can be misleading when predictors are causally related; the paper seeks causally valid univariate explanations of black-box models.

Approach

Use an auxiliary explanatory causal model (ECM) that embeds the black-box prediction as a node, and compute counterfactual curves of the prediction under interventions on a predictor (total, partially controlled, natural direct, natural indirect), showing PDP/ICE as special cases corresponding to natural direct effects.

Key results

• On a simulated nonlinear mediation example, CDPs show the linear black-box's total dependence on X looks cubic while the true relationship is quadratic, with the cubic component attributable to the indirect path through M.
• On the Breast Cancer Wisconsin data, the PC algorithm yields an undirected edge between Cell Size and Cell Shape, and the takeaway about Cell Shape's effect on tumor class is sensitive to the edge orientation, especially for total dependence.
• The paper shows that the PDP/ICE coincide with the natural direct dependence plot under explanatory data drawn from the causal model (Proposition on partial dependence as natural direct effects).

Measured effects reported in the paper

Surprising findings

• The NDDP (equivalent to PDP) can show a trend that is practically reversed relative to the total dependence for the PKA example.

Limitations

• CDPs depend on the assumed or learned auxiliary causal model; mismatch with the true DGP can mislead.
• If the black-box fails to fit the true DGP, explanations about the DGP will also fail.
• Causal conclusions require untestable causal assumptions.

Experimental setup

Datasets: Breast Cancer Wisconsin (Original) synthetic mediation simulation motivating salary example (synthetic)

Benchmarks: Breast Cancer Wisconsin (Original)

Metrics: total dependence natural direct dependence natural indirect dependence partially controlled dependence

Baselines: Partial Dependence Plot Individual Conditional Expectation

Builds on: Partial Dependence Plot Individual Conditional Expectation structural causal models PC algorithm DoWhy

Closest papers by meaning

• Identifying Macro Causal Effects in C-DMGs over DMGs (NeurIPS 2025 · similarity 0.85)
• Local Identifying Causal Relations in the Presence of Latent Variables (ICML 2025 · similarity 0.84)
• A Neural Framework for Generalized Causal Sensitivity Analysis (ICLR 2024 · similarity 0.84)
• The third pillar of causal analysis? A measurement perspective on causal representations (NeurIPS 2025 · similarity 0.84)
• Causality can systematically address the monsters under the bench(marks) (NeurIPS 2026 · similarity 0.84)
• Unifying Causal Representation Learning with the Invariance Principle (ICLR 2025 · similarity 0.84)