Interpretability Frameworks for Continuous Value Regression

34 concepts30 papers

Most explainable AI (XAI) research has historically focused on classification: the task of determining why a model chose "Label A" over "Label B." In these scenarios, interpretability is anchored by the decision boundary—the threshold where a prediction flips from one category to another. However, as noted by David Baehrens et al. in How to Explain Individual Classification Decisions, these frameworks do not naturally generalize to regression.

In regression, the output is a continuous scalar—a house price, a steering angle, or a disease severity score. There is no discrete boundary to cross. This "regression explanation gap" requires a fundamental shift in how we define what an explanation is and how we calculate the influence of input features on a numerical range.

Redefining Feature Attribution for Scalars

Feature attribution methods like SHAP or Layerwise Relevance Propagation (LRP) were originally designed to explain class probabilities. To make them work for regression, researchers must redefine how "relevance" flows through a network toward a single numerical value.

Ashkan Khakzar et al. addressed this in Explaining COVID-19 and Thoracic Pathology Model Predictions by Identifying Informative Input Features by developing Regression Information Bottleneck Attribution (IBA). This method identifies which regions of an input (such as a chest X-ray) influence a continuous severity score rather than a binary diagnosis. Similarly, Lingbai Kong introduced Regression Relevance Propagation (RRP) in CausalFormer: An Interpretable Transformer for Temporal Causal Discovery, allowing for the quantification of how specific time-series variables contribute to a future continuous state.

The core challenge here is the reference point problem. As Tülay Adali et al. argue in Interpretability, Reproducibility, and Replicability [From the Guest Editors], regression requires a unique approach to choosing reference points—baselines against which a numerical change can be measured—which is often overlooked when applying classification-centric methods.

Counterfactuals without Decision Boundaries

Counterfactual explanations usually answer: "What is the smallest change needed to change the outcome?" In classification, the target is clear: the other class. In regression, the target is an arbitrary point on a continuous line.

Klaus Broelemann et al. tackle this in Counterfactual Explanation for Regression via Disentanglement in Latent Space by disentangling label-relevant features in a latent space. Instead of searching for a boundary, their method performs linear interpolation to find realistic paths to a target numerical outcome. This is critical because, as Patrick Altmeyer et al. note in Explaining Black-Box Models through Counterfactuals, providing actionable recourse in numerical domains (like loan amounts or physical trajectories) requires methods that can handle the lack of discrete target states.

Symbolic Rules and Case-Based Reasoning

Another path to interpretability is replacing the black box with a transparent proxy, such as a set of IF-THEN rules. However, defining a "same decision" for a rule is difficult when the output varies continuously.

The Complexity of Continuous Interpretability

As these methods scale to more complex architectures, new theoretical hurdles emerge. Shahaf Bassan et al. demonstrate in What makes an Ensemble (Un) Interpretable? that ensembles are exponentially harder to interpret than single models, a challenge that persists and potentially intensifies in continuous output spaces. Furthermore, Marcelo Arenas et al. point out in The Tractability of SHAP-Score-Based Explanations for Classification over Deterministic and Decomposable Boolean Circuits that the structural properties allowing for efficient explanation in Boolean logic do not easily generalize to models outputting continuous values.

Whether through graph-based explanations for molecular properties (Jiaxing Zhang et al. in RegExplainer) or instance-level attribution in Multiple Instance Learning (Joseph Early et al. in Model Agnostic Interpretability for Multiple Instance Learning), the research direction is clear: moving beyond the "flip" of a classification label to explain the "flow" of continuous numerical prediction.

Key Foundational Work

How to Explain Individual Classification Decisions (2009) — contributed 1 concept elements to this direction.
Rule Extraction with Guaranteed Fidelity (2014) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 29 papers
  2. Version 1 · 2026-10-04 · 42 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_continuous_value_attribution_788dedce,
  title = {continuous value attribution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-value-attribution/rdr_788dedceadb8424ea2198988f64187f5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_continuous_value_attribution_788dedce_v2,
  title = {continuous value attribution},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/continuous-value-attribution/rdr_788dedceadb8424ea2198988f64187f5/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}

See something wrong? Report a mistake