Comparables XAI: Faithful Example-based AI Explanations with Counterfactual Trace Adjustments

arXiv:2602.13784 · cs.HC, cs.AI · Submitted 2026-02-14 · 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 "Comparables XAI: Faithful Example-based AI Explanations with Counterfactual Trace Adjustments".

Jane: The paper was written by Yifan Zhang, Tianle Ren, Fei Wang and Brian Y Lim from Department of Computer Science, National University of Singapore and National University of Singapore, Singapore, Singapore and University of Singapore, University of Singapore, United States of America (implied by the context/location).

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

Paper discussion segment 2: Tom: So, we’ve established that simply comparing is intuitive, but it lacks depth when the gap between two real examples is huge. The paper's summary shows they bridge this gap by modeling a specific counterfactual journey.

Jane: They are essentially creating a hypothetical path for the "Subject" case, showing how it relates to its nearest "Comparables" by applying systematic adjustments.

Lu: This feels like moving beyond just establishing similarity; it’s about respecting the dynamics of the decision surface itself, which is a massive theoretical leap.

Meng: I'm curious about how they define that path—they aren't just guessing; they are simulating a specific, measurable transformation from the target case to look more like a reference.

Lalam: The goal is making this complex process understandable, so that the explanation becomes transparent and convincing for everyone involved in trusting AI.

Paper discussion segment 3: Tom: We know that simple linear adjustments are too basic, but the authors didn't stop there. They introduced "Comparables with Trace Adjustments" to make this much more accurate.

Jane: This is where the method moves beyond just making one big adjustment; it traces a path of counterfactual changes incrementally, one attribute at a time.

Lu: I think this incremental approach is brilliant for my theoretical model because it allows us to view the non-linear decision surface as a piecewise linear function that traverses the feature space between those two points. It respects the curvature of AI’s logic instead of ignoring it.

Meng: From an engineering standpoint, modeling this trace using discrete segments makes sense for computation, but we have to ensure we can read it without getting overwhelmed by small changes in a large dataset.

Lalam: The idea is making the complex accessible, so that the users can examine small, meaningful changes between each step and build trust through incremental understanding.

Tom: They also formalized this process using five specific criteria—like sparsity and disjointness—to help select the best way to trace that path.

Paper discussion segment 4: Tom: The authors didn't just run a single simulation; they conducted extensive modeling and user studies to prove "Comparables XAI" is genuinely better than baseline methods.

Jane: In the modeling study, they found that Trace-adjusted Comparables consistently outperformed others, achieving the highest XAI faithfulness and precision across all methods tested.

Lu: The fact that this approach yielded the narrowest uncertainty bounds is a huge win for me, because it means we are giving users a much tighter range of expected outcomes than previous methods allowed us to estimate.

Meng: I’m really impressed by the sensitivity analysis they performed on those five criteria; it showed that you can precisely control the trade-off between how readable an explanation is and how accurate it is, which is essential for making this work in a real product.

Lalam: This truly demonstrates that we can achieve a balance—it's not just about having an explanation; it’s about having an *optimal* explanation based on human cognitive load.

Tom: So, the modeling study confirms the technique is technically sound and accurate, but how does this translate to real people actually using it?

Jane: That’s where they moved into their formative user study to see if users understood or struggled with these "Comparables with Trace Adjustments" versus other methods.

Conclusion: Tom: We've covered the theory, the mechanics, and the user testing for "Comparables XAI: Faithful Example-based AI Explanations with Counterfactual Trace Adjustments," and it's clear this is a major step forward in making AI understandable.

Jane: I think it’s incredibly encouraging to see that users found trace adjustments to be more reliably consistent and easier to interpret than the other methods we looked at.

Lu: The way the authors have structured these counterfactual traces—step by step, respecting the non-linearity of a piecewise function—is a genuine theoretical advancement in how we model human intuition alongside AI logic.

Meng: My final take is that this approach can be scales to any valuation task, not just real estate; it provides a practical framework for explaining anything from company value to energy consumption prediction.

Lalam: I hope that this is the foundation for better user trust across all forms of developing AI, providing a path forward where we can say goodbye and move on.

Tom: That’s right, and wrapping up our discussion on "Comparables XAI: Faithful Example-based AI Explanations with Counterfactual Trace Adjustments" for today, we hope you found this conversation as insightful as we did.

Jane: We'll be back next time to look at some fascinating new papers!

Department of Computer Science, National University of Singapore · National University of Singapore, Singapore, Singapore · University of Singapore, University of Singapore, United States of America (implied by the context/location)

cs.HC, cs.AI

Submitted: 2026-02-14

Updated: 2026-09-04

Comments: Accepted by CHI 2026

DOI: 10.1145/3772318.3791041

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

Importance score: 84/100

The gist: AI explainability (XAI) is a critical area of research intended to foster user understanding and trust in AI systems; however, many existing XAI methods are "too technical" for non-technical users.

Key concepts

Counterfactual Trace Adjustments
This method simulates a hypothetical path between two data points by applying systematic adjustments incrementally, one attribute at a time. It avoids large jumps and respects the complex, non-linear dynamics of AI decision surfaces to make the process measurable.
Comparables XAI
This is the framework that uses example-based comparisons but adds depth by modeling the counterfactual journey. It provides transparent explanations showing how a specific case relates to its neighbors through measurable, step-by-step transformations.
XAI Faithfulness and Precision
These metrics measure how accurately and reliably an AI explanation reflects the actual underlying model logic. The trace-adjusted method was found to achieve higher levels of both faithfulness and precision compared to other existing methods.

Terminology

Summary

AI explainability (XAI) is a critical area of research intended to foster user understanding and trust in AI systems; however, many existing XAI methods are too technical for non-technical users. While example-based explanations are intuitive, they often fail to maintain faithfulness when examples differ significantly from the target case. This work addresses this gap by proposing Comparables XAI, a framework that uses relatable example-based explanations combined with counterfactual adjustments to provide a more accurate and transparent justification for AI decisions.

How it works

The core idea is inspired by real estate valuation, where appraisers estimate the price of a Subject property by applying systematic adjustments to the known prices of similar properties (Comparables). Comparables XAI extends this concept to AI decision-making. Instead of simply presenting raw examples, it uses Trace adjustments to provide a counterfactual path that incrementally adjusts from each Comparable toward the Subject. This process is modeled as a piecewise linear function that traverses the feature space, allowing users to examine small changes between each counterfactual step, thereby limiting cognitive load and improving trust.

The Technical Approach

The paper evaluates four methods: 1) Example-Based Explanations (weighted average), 2) Comparables with Linear Regression (fitting a straight trend line), 3) Comparables with Linear Adjustments (using local linear models to approximate counterfactual examples), and the proposed method, 4) Comparables with Trace Adjustments. The latter is designed to overcome the limitations of linear adjustments, which assumes a straight path between the Comparable and the Subject. Trace Adjustments models this transition as a piecewise function defined by several key desiderata:

  • Sparsity (LS): To reduce cognitive load by minimizing the number of terms in the explanation.

  • Disjointness (LD): To ensure each attribute changes only once along the trace.

  • Monotonicity (LM): To maintain a consistent direction of change in values and attributes.

  • Evenness (LE): To ensure stability by penalizing the variance of consecutive label differences.

The technical implementation uses gradient descent to train this piecewise function, ensuring the path faithfully follow[s] the nonlinear shape underlying function of the AI model.

Evaluation and Results

A modeling study was conducted across five application domains (House Price, Salary, Energy Consumption, Drug Sensitivity Analysis, and Crop Yield Estimation). The results consistently showed that Comparables with Trace Adjustments achieved superior performance compared to baseline methods: the highest XAI faithfulness and precision alongside the narrowest uncertainty bounds.

In a formative user study involving 25 participants from diverse backgrounds, users found that the trace-based approach was perceived as more tightly consistent and usably interpretable, especially appreciating the step-by-step detail. The summative user study confirmed these findings, showing that Trace Adjustments led to significantly lower decision mean error and narrower credible intervals than all other XAI types.

Conclusion

Comparables XAI provides a new analytical paradigm for example-based explanations, allowing users to analytically compare examples in a way that is both intuitive and highly accurate. This approach successfully bridges the gap between the simplicity of example-based explanations and the complexity of non-linear AI decision surfaces, fostering greater confidence in AI-assisted decision-making.

Improvements for AI systems

Based on the findings of this research, I have identified several critical architectural and methodological improvements that must be implemented to elevate current AI systems from merely generating predictions to providing truly trustworthy and intuitive explanations.

The core improvement is not just adding a counterfactual, but implementing a Trace Adjustment mechanism governed by specific constraints (Desiderata) that ensure the explanation aligns with human cognitive models.

  1. Transition from Global Linear Interpolation to Piecewise Trace Modeling:

  • Current System Failure: Most XAI methods assume a single, linear path between the Subject and Comparables, leading to inaccurate results when the underlying AI model (f) is non-linear (as shown in Figure 2c).

  • The Improvement: Implement a Piecewise Linear Function framework. The system must treat the path from Comparable (x c) to Subject (x s) not as one vector, but as a sequence of N segments (tau). Each segment is defined by a localized linear equation: y(x) = w tau T x + b tau.

  • The Outcome: This allows the system to model the complex, non-linear decision surface of the AI model in small, manageable steps, ensuring that as a user moves incrementally from one attribute change to the next, they are following a path that faithfully approximates f(x).

  1. Integrating Desiderata Constraints into the Trace Optimization:

  • The trace must not be arbitrary. The system must incorporate five specific constraints (Desiderata) during the calculation of these incremental steps (tau) to optimize for human perception and trust:

  • Sparsity (L 1 loss): Minimize the total number of attribute changes across the trace, reducing cognitive load for users.

  • Disjointness (T 1): Ensure each attribute changes only once along the entire trace, preventing confusing or redundant adjustments.

  • Monotonicity (T 2): Constrain the direction of change in both attributes and their corresponding label values (x and y) to be consistently increasing or decreasing, reducing surprise.

  • Evenness (L E): Penalize the variance of consecutive value changes (y), ensuring a stable, predictable rate of adjustment.

The implementation of Comparables XAI enables the AI system to perform functions that current methods cannot achieve:

  1. Provide Faithfully Consistent Explanations: The improved system can generate an explanation where every single step—from the initial Comparable to the final Subject—is a faithful approximation of how the original, complex AI model would have arrived at its prediction. It moves beyond good enough approximations to model-aligned explanations.

  2. Facilitate Incremental Cognitive Understanding: Instead of presenting one large, abstract adjustment (e.g., +65k ), the system provides a step-by-step counterfactual trace. A user can observe exactly how the price changes as a single attribute is adjusted (e.g., If Living Area decreases by 10 sqft, the price drops by X amount), allowing them to verify that specific, local influence.

  3. Increase User Confidence and Trust: By optimizing for Desiderata (especially Sparsity and Monotonicity), the system provides an explanation that is not only accurate but also intuitive. Users will no longer be confused by spurious or inconsistent adjustments, leading to higher user accuracy in decision-making (as demonstrated in the summative study).

  4. Scale to Multi-Attribute Complexity: The system can successfully handle complex domains (like real estate valuation or drug sensitivity) where multiple attributes interact non-linearly, allowing users to understand how changes across different dimensions contribute simultaneously without the explanation becoming an opaque black box.

Abstract

Explaining with examples is an intuitive way to justify AI decisions. However, it is challenging to understand how a decision value should change relative to the examples with many features differing by large amounts. We draw from real estate valuation that uses Comparables-examples with known values for comparison. Estimates are made more accurate by hypothetically adjusting the attributes of each Comparable and correspondingly changing the value based on factors. We propose Comparables XAI for relatable example-based explanations of AI with Trace adjustments that trace counterfactual changes from each Comparable to the Subject, one attribute at a time, monotonically along the AI feature space. In modelling and user studies, Trace-adjusted Comparables achieved the highest XAI faithfulness and precision, user accuracy, and narrowest uncertainty bounds compared to linear regression, linearly adjusted Comparables, or unadjusted Comparables. This work contributes a new analytical basis for using example-based explanations to improve user understanding of AI decisions.

Sources

Related papers