A Computationally Feasible Framework for Causal Probabilistic Explanation

arXiv:2609.04177 · cs.AI · Submitted 2026-09-03 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Next we'll be talking about the paper "A Computationally Feasible Framework for Causal Probabilistic Explanation".

Jane: The paper was written by Rafal Urbaniak, Sam Witty, Andy Zane, Emily Bunnapradist, Daniel Waxman et al. from Basis Research Institute, University of Massachusetts Amherst, Massachusetts Institute of Technology, University of California, Los Angeles, Sorbus AI, HouseIQ.

Tom: Stay tuned as we take you through the paper and discuss its implications.

Title: Tom: We're looking at this paper "A Computationally Feasible Framework for Causal Probabilistic Explanation," and right away, it tackles a massive problem in AI attribution research. The authors are clearly pointing out that much of the existing work is either too complex to run on massive datasets or it ignores the actual causal story that happened.

Jane: That's a huge point, Tom; they acknowledge how hard it is to compute true causality for large-scale systems. They aren't just trying to make a tool; they are designing a coherent framework for what we should be expecting from complex AI behavior.

Lu: This sounds like the foundational work that’s necessary for moving beyond simple discrete decision trees, right? The way they incorporate probabilistic machinery suggests they are ready to handle continuous variables and vectors, which is vital for modern AI systems.

Meng: I'm looking at this from a practical standpoint, and having feasibility is critical because it allows the us to actually run these kinds models on real-world things like an automated valuation system without the entire system being throttled by trying to calculate every single possible causal path.

Lalam: It's encouraging to see them focusing on "Causal Probabilistic Explanation" because it implies they are taking the probabilistic nature of reality into account, not just some idealized deterministic model where cause and effect is a simple one-to-one line.

Tom: So, while the authors have established this clear goal—bridging theory and practice—we need to look at how this framework operates when we see their results in the next segment.

Paper discussion segment 2: Jane: The abstract mentions that they evaluate PCI on synthetic examples, but the real-world model of an Automated Valuation Model (AVM) trained on millions of data points is just as important. This shows they aren're not limiting this framework to academic puzzles.

Tom: And the results are impressive because PCI matches the verdicts from actual causality on those canonical archetypes, which is a big deal since it validates their approach against established theory. But I think there's something subtler than just matching outcomes—they are proving its ability to handle situations where multiple causes exist without one preempting the other.

Lu: That’s the concept of overdetermination, and being able to solve that means a deep understanding how all those pathways contribute simultaneously is necessary. It's not just about finding *a* cause; it's about mapping the entire causal landscape.

Meng: From an implementation standpoint, I think the fact that they are using Monte Carlo sampling means we can estimate these complex contributions in a predictable amount of time, even on massive datasets like twenty-five nodes and ninety-nine edges. That's exactly what my team needs for large-scale systems.

Lalam: The implication for us is that when AI shows multiple possible causes at once, we won't have to guess which one is the "right" answer; instead, we get a graded score reflecting how much each contribute to the overall uncertainty.

Tom: It seems like they’ve successfully demonstrated that this method can manage complexity while keeping the math manageable, so let's move on to look at what specific structural improvements they built into PCI.

Paper discussion segment 3: Jane: The authors identified that traditional attribution methods like SHAP are often correlation-based, but they aren't context-sensitive—meaning they don't care about the specific causal story that led to a result. PCI addresses this by incorporating what they call a "witness mechanism."

Tom: That witness concept is really smart because it lets you hold certain parts of the causal story fixed at their real-world values. It’s like saying, "I know the system was working in this specific way," which makes it much more grounded than just looking at correlations.

Lu: The way they integrate a "variable selection distribution" to choose which suspects to examine is also a major refinement. It prevents us from getting overwhelmed by trying every single variable, focusing our attention on the most plausible causal paths that matter.

Meng: I think the combination of Monte Carlo sampling with this structured choice of suspects and witnesses means that the operational cost scales well across different scenarios, which is ideal for my systems when we are not hitting those exponential limits.

Lalam: It’s a powerful way to show intent by using context; instead of just seeing a feature's value, we are seeing its contribution *within* the specific causal context of the event. That makes AI explanations much more human-readable for people who need to understand the why behind a decision.

Tom: It's clear they have engineered several specific parts—the sampling, the context, and the structure—to fix these known issues in old methods; let's see how this works when we look at their empirical tests.

Conclusion: Jane: So, to conclude our discussion of "A Computationally Feasible Framework for Causal Probabilistic Explanation," we’ve seen that this is a major leap forward in AI attribution. It gives us the ability to understand *why* an outcome occurred while remaining practical enough to compute at scale.

Tom: It's truly a unified, graded measure that moves past binary "yes/no" verdicts, which is exactly what we needed when looking at complex systems where causes are intertwined. I think it allows for a level of nuance that was simply impossible before this work.

Lu: This framework provides the necessary structure to handle the complex interplay between many causal factors and continuous outcomes, which is where the theoretical physics of causality usually breaks down in current AI tools.

Meng: From a practical standpoint, I think this offers a clear path toward genuine transparency in real-world deployment of these large models without compromising computational efficiency. We have a way to get the right answer at scale.

Lalam: My final thought is that it allows us to build AI that isn't just predicting things but understanding its own internal logic and be accountable for its decisions, ensuring we are truly building understandable agents.

Tom: It has been a fascinating look at "A Computationally Feasible Framework for Causal Probabilistic Explanation," and I think the work is doing a lot of heavy lifting in the right directions for this field.

Basis Research Institute, University of Massachusetts Amherst, Massachusetts Institute of Technology, University of California, Los Angeles, Sorbus AI, HouseIQ

cs.AI

Submitted: 2026-09-03

Updated: 2026-09-03

Code: https://github.com/rfl-urbaniak/explainable_paper

Project page: https://rfl-urbaniak.github.io/explainable_paper

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 91/100

The gist: This paper introduces a novel framework for causal probabilistic explanation, demonstrating how to rigorously quantify causal claims in complex machine learning models.

Key concepts

Causal Probabilistic Explanation (PCI)
A framework designed for AI attribution that provides a graded measure of cause and effect. Unlike simple correlation, it accounts for the probabilistic nature of reality and multiple simultaneous causes, offering a nuanced explanation of why an outcome occurred.
Overdetermination
A concept where multiple causes exist simultaneously contribute to an outcome. The framework is designed to map this entire causal landscape, allowing users to understand how all contributing pathways affect the final result.
Witness Mechanism
A structural improvement in PCI that allows certain parts of the causal story to be held at their real-world values. This grounds the explanation in reality, making it more accurate than simply looking at correlations.

Terminology

Summary

This paper introduces a novel framework for causal probabilistic explanation, demonstrating how to rigorously quantify causal claims in complex machine learning models. By developing PCI (Probabilistic Causal Inference), the authors provide a finer-grained verdict than the binary 'actual cause / not' that classical AC produces, allowing researchers to distinguish between robust causes and context-sensitive effects across counterfactual scenarios. This capability is crucial for moving causal inference from controlled academic settings to real-world, high-stakes automated valuation models (AVMs).

Causal Inference Frameworks and Benchmarks

The framework employs PCI to evaluate causal influence by analyzing how policy changes affect outcomes across multiple factual worlds. The authors show that PCI can recover complex intuitions that simpler methods miss. For instance, on the dynamical SIR setting, PCI successfully distinguishes between policies where lockdown is the proximate cause of excessive overshoot and masking only a contextsensitive one, even when these policies appear similar under traditional but-for analysis. Furthermore, the methodology provides distinct decompositions:

  • The necessity-vs-sufficiency decomposition distinguishes suspects whose isolation effects are similar but whose removal-under-context effects diverge.

  • The contexts experiment further distinguishes robust from context-sensitive causes.

Application to Automated Valuation Models (AVMs)

A significant advancement is the application of PCI to a deployed, real-world AVM for residential housing where neither ground truth nor exact AC enumeration is available. The authors establish that the method is computationally viable at production scale. The core components of this system include:

  • Model Structure: The outcome variable y (the log-epsilon-standardized sale price) is modeled by a sparse variational Gaussian Process (SVGP).

  • Causal Kernels: Variables are drawn from causal kernels, which are deep neural networks optimizing distributions that can be either truncated normal or categorical.

  • Feasibility Claim: The process is deemed computationally tractable at production scale, requiring an estimated 60 minutes to estimate expected values for a batch of 50 data points using Monte Carlo sampling.

Comparison with Attribution Methods (SHAP vs. PCI)

The authors provide a qualitative comparison between PCI attributions and established methods like SHAP scores on the AVM prototype. They observe that the attribution mechanisms differ significantly due to how they handle variable dependencies:

  • SHAP's observational marginalisation allocates weight differently from intervention-and-witness-based PCI scores.

  • On a deep DAG, this difference manifests as SHAP's tendency to assign disproportionate attribution mass to nodes that act like summary statistics of the outcome, whereas PCI distributes attribution more evenly across the upstream causal structure.

Technical Implementation Details

The underlying causal assumptions are formalized using a Causal Directed Acyclic Graph (DAG) G, which expresses relationships between variables. The model construction involves:

  1. Defining the unit of analysis (e.g., a residential property transaction).

  2. Manually writing the DAG in collaboration with domain experts, followed by refinement to break implied conditional independencies not upheld by the data.

  3. Jointly optimizing the SVGP outcome model with a linear embedding (eta) of categorical variables into the continuous space, which is optimized alongside the GP parameters (phi mu, phi kappa).

Improvements for AI systems

The core scientific contribution detailed in this paper is not merely a new attribution score, but an advanced framework for contextual and dynamical causal inference that operates within complex, high-dimensional, non-linear models (like AVMs).

Based on this methodology, I propose the following specific improvements to existing AI systems.

Improvement: Integrate a dedicated module into standard XAI pipelines that moves beyond simple counterfactual perturbations or observational marginalization (like SHAP). This module must explicitly model the conditional dependence of feature influence on the state of other features being held constant (witness sets).

Mechanism Detail:

  1. Witness-Conditioning Layer: The CCAM must accept a set of designated witness variables (W). When calculating the attribution score for a suspect variable (X i), it must calculate the impact not just under its own intervention, but specifically conditioned on the values assigned to W.

  2. Contextual Score Decomposition: The system must decompose the total attribution into three quantifiable components:

  • Necessity Score (Baseline): The impact when X i is removed/set to a baseline, holding all other variables constant.

  • Sufficiency Score (Contextual Uplift): The incremental gain in prediction when X i is introduced, relative to the necessity score, given the witness context W.

  • Robustness Score: A statistical measure quantifying how sharply the Necessity and Sufficiency scores change across slightly perturbed witness settings (W plus or minus epsilon). A low Robustness Score indicates a context-sensitive cause.

Improved System Capability:

The resulting AI system can distinguish between:

  • True Causal Drivers: Features whose influence remains high and stable regardless of the context (high Necessity/Sufficiency, high Robustness).

  • Context-Dependent Triggers: Features that only exert influence when a specific set of conditions are met (low Robustness, but high score in specific W contexts). This is critical for risk assessment where environmental or operational states change.

Abstract

Explaining why a specific outcome occurred, and which inputs deserve the blame or credit, is central to philosophical, scientific, and policy analysis. Existing tools split into two camps. The theory of actual causality (AC) gives principled verdicts, but only for toy-sized models, because computing them requires enumerating counterfactual scenarios. Scalable attribution methods like SHAP (or even causal SHAP) at least partially ignore the causal structure that generated the data, and can give answers that conflict with a careful causal analysis. We close this gap with Probabilistic Causal Impact (PCI). PCI builds on actual causality and on Pearl's notions of probability of necessity and sufficiency, but recasts the question of explainability as an estimation problem on a probabilistic causal model that is easily approximated via Monte Carlo. By specifying a distribution over "candidate explanations," a distribution over counterfactual values, and a scoring function, PCI provides tractable, causally grounded, graded explanations, generalizing AC and Pearl's probability of causation as degenerate cases. We evaluate PCI in synthetic and real-world examples, spanning consistency checks with AC, scaling experiments, complex continuous-valued dynamical systems, and a real-world deployed causal machine learning model trained on millions of datapoints.

Sources

Related papers