Causal Potential Scoring Frameworks

• Causal potential scoring is a quantitative framework that assigns scores to causal relationships by integrating prior beliefs and data likelihood across Bayesian networks, time series, and machine learning models.
• It leverages methods like IPFP, zero-Hessian tests, and neural interventions to enhance structural discovery, model interpretability, and robust estimator selection.
• Empirical results demonstrate improvements such as a 10–30% reduction in structural errors and better alignment with human judgments, though challenges remain with prior specification and sample size.

Causal potential scoring refers to a class of quantitative frameworks for assigning, estimating, or leveraging scores that reflect the strength or structure of causal relationships within observed data, candidate models, or machine learning outcomes. Approaches span Bayesian network structure learning, time series causal effect estimation, neural model interpretability, automated dialogue evaluation, LLM-based edge prediction, score-based causal discovery, and principled model selection in causal inference pipelines. Core motives include enabling the incorporation of prior causal knowledge, the automation of causal graph construction, the identification of causally significant features or regions, and the data-driven validation of relevance or effectiveness in complex settings.

1. Causal Potential Scoring in Bayesian Network Structure Search

In Bayesian network structure learning, “causal potential” scoring formalizes how subjective beliefs about pairwise causal (or associative) relations—termed "path beliefs"—are converted into a prior over directed acyclic graphs (DAGs), then combined with data likelihoods to guide search (Borboudakis et al., 2012).

• For each (X, Y) pair with path belief, introduce a variable $r_{X,Y}$ with 4 states: “$X\to Y$”, “$Y\to X$”, “common-ancestor”, “independent”.
• User beliefs take the form of marginal probabilities $$\Pi_{r_{X,Y}} = (\pi_{X\to Y}, \pi_{Y\to X}, \pi_{CA}, \pi_\perp)$$ partitioning all d-connecting relationships.
• The scoring process constructs a global joint $J$ over all path variables, subject to marginal constraints and coherence (no cycles).
  • If marginals are coherent, $J$ solves a constrained minimum-$D_\mathrm{KL}$ divergence problem to the uninformative configuration prior $U$ (number of DAGs per configuration) using IPFP.
  • For incoherent sets, GEMA (a relaxed IPFP) is used.
• Each DAG $G$ induces an equivalence class/configuration $C_G$; the prior is assigned as $X\to Y$0, and the prior score is $X\to Y$1.
• The total score $X\to Y$2 is maximized via greedy local search, augmented by a swap-equivalent operator that explores all Markov-equivalent DAGs to optimize $X\to Y$3.
• The method is equivariant under Markov equivalence and becomes uniform if path beliefs are uninformative.

Empirical results show path-belief driven scoring reduces structural Hamming distance by 10–30% over uniform priors in moderate sample regimes, supporting both skeleton and orientation learning (Borboudakis et al., 2012).

2. Score Matching and Potential in Causal Discovery

In score-based causal discovery, the “score function” $X\to Y$4 is interpreted as a causal potential, revealing conditional independence and causal ordering structure from data (Montagna et al., 2024).

• Proposition 1: The mixed second partial derivative $X\to Y$5 indicates $X\to Y$6 and $X\to Y$7 are m-separated in the PAG on $X\to Y$8.
• In additive noise models (ANMs), direct edges are detected by regressing node-wise scores on residuals $X\to Y$9; zero mean square error identifies sinks.
• Extensions include conditions for latent variable settings: vanishing score-residual MSE characterizes direct causes in a marginal MAG.
• The AdaScore algorithm alternates between skeleton discovery (using zero-Hessian tests) and orientation (via score-residual MSE minimization), supporting both linear and nonlinear, as well as latent variable, SEMs.
• Score matching estimators (e.g., Stein gradient) are used to estimate the score and its Jacobian/nonparametrically.

Synthetic benchmarks confirm the competitive performance of this potential-based scoring under varied structure and hidden variable models (Montagna et al., 2024).

3. Causal Potential Scoring in Temporal and Experimental Causal Inference

In the Potential System (PS) framework for time series, the dynamic causal effect $Y\to X$0 operationalizes the time-lagged causal potential of interventions (Carlson et al., 20 Mar 2026).

• PS models assignments $Y\to X$1 and responses $Y\to X$2 within a nonparametric, counterfactual syntax.
• Causal potential scores correspond to these dynamic causal effects—ATE, CATE, FTE, and CFTE—each defined as averages over different sources of randomness and conditioning.
• Identification relies on sequential ignorability and overlap conditions for assignment mechanisms (SAM.BSU, SAM.BSR, etc.).
• Estimation employs kernel regression, local linear regression, or local projections, and impulse-response analysis in SVAR.
• Scoring consists of reporting $Y\to X$3 trajectories, confidence bands, and design-based inference.

This formalizes the direct mapping between dynamic causal potential and nonparametric effect estimation in time series with strong identification rationale (Carlson et al., 20 Mar 2026).

4. Causal Potential Scores in Model Explanations and Machine Learning Metrics

Quantifying the causal contribution of input regions or features is addressed by the Causal Explanation Score (CaES) in deep medical image classification (Villegas-Jimenez et al., 2023) and by CausalScore in dialogue evaluation (Feng et al., 2024).

• CaES: Interventions replace image regions (object-only vs. context-only). Feature activation changes are normalized per feature, transformed with a sigmoid, and aggregated to give class-level scores. This validates whether a region causally drives the classifier’s output, extending beyond traditional saliency.
  • Experiments show Grad-CAM–derived masks yield CaES indistinguishable from human-segmented regions, and that object-only CaES is substantially larger, matching domain priors (Villegas-Jimenez et al., 2023).
• CausalScore: In open-domain dialogue evaluation, the causal potential between context utterances and response is computed via learned tests of unconditional and conditional dependence, interpreted as proxies for causal strength.
  • The final score aggregates both unconditional and conditional dependence metrics to align closely with human judgments of relevance, outperforming reference-based metrics (Feng et al., 2024).

Both approaches systematically convert interventions or classifier-based dependence tests into a quantitative measure of causal responsibility, bridging black-box explainability with empirically validated causal scoring.

5. Automated and Data-Driven Causal Potential Scoring with LLMs and AutoML

Applications have extended causal potential scoring to automated graph construction with LLMs and out-of-sample estimator selection in AutoML pipelines (Long et al., 2023, Kraev et al., 2022).

• LLM-based edge scoring: “Causal potential” is operationalized by comparing LLM likelihoods for causal vs non-causal edge statements (e.g., “X increases the risk of Y” vs. its negation). The log-prob gap is interpreted as a causal-potential score. Results indicate substantial prompt sensitivity and no acyclicity enforcement, necessitating expert-in-the-loop validation (Long et al., 2023).
• Model selection via causal scoring: In AutoCausality, true individual treatment effects are unobservable; thus, scoring metrics like Normalized ERUPT (for CATE) and energy distance (for IV) are adopted as proxies for causal potential to enable out-of-sample model selection.
  • These metrics, grounded in policy-reward or balance after effect adjustment, allow reliable estimator ranking and tuning in synthetic and real-world settings, outperforming naive baselines (Kraev et al., 2022).

These results demonstrate the portability of “causal potential” scoring to contexts where explicit structure learning or model ranking is needed, leveraging statistical proxies and scalable estimation.

6. Practical Implications, Experimental Results, and Limitations

Causal potential scoring provides a principled bridge between subjective prior knowledge, empirical causal effect estimation, automated discovery, and model selection. Empirical evaluations (Bayesian networks, ANMs, time series, neural inductive tasks) consistently show:

• Substantial improvements in recovery of true network structure or effect accuracy when path or feature-level causal knowledge is formally incorporated (Borboudakis et al., 2012, Villegas-Jimenez et al., 2023).
• Robustness in adversarial or small-sample regimes, with graceful degradation under mis-specified priors (Borboudakis et al., 2012).
• In dialogue and explainability evaluation, direct dependence/causal-potential scores align better with relevance than reference-based metrics or black-box importance (Feng et al., 2024).
• AutoML pipelines for causal effect estimation benefit from principled out-of-sample scoring proxies rooted in causal potential or balance principles (Kraev et al., 2022).

Limitations include:

• Dependence on accurate prior marginal probabilities or sufficient sample size for coherent joint estimation in Bayesian networks (Borboudakis et al., 2012).
• Model class limitations for score-based discovery; performance can deteriorate when data strongly violate the faithfulness or additive noise assumptions (Montagna et al., 2024).
• Prompting brittleness and context sensitivity in LLM-based scoring (Long et al., 2023).
• Need for sufficient validation samples and reliable propensity/instrument estimation in policy-value or balance-based scoring metrics (Kraev et al., 2022).

In summary, causal potential scoring underpins a diverse array of state-of-the-art statistical, machine learning, and automated discovery methodologies, providing a rigorous, theoretically justified foundation for encoding, estimating, and validating causal effects and dependencies across data types and domains.