Delta-XAI: Explaining Prediction Changes

• Delta-XAI is a paradigm in explainable AI dedicated to elucidating changes in predictions rather than static outputs.
• It leverages methods such as DeltaSHAP, SWING, DeltaXplainer, δ-XAI, and DXAI to provide directional, magnitude-sensitive, and context-aware attributions.
• The framework improves decision-making in clinical monitoring, financial analysis, and model lifecycle management by addressing model drift and temporal evolution.

Delta-XAI refers to a paradigm and the associated family of methods in explainable artificial intelligence (XAI) dedicated to attributing and visualizing the differences or evolutions in predictions, rather than static model outputs. These methods target scenarios where stakeholders require transparency into why a model's decision or risk score has changed over time, across input states, or between different model versions. Unlike conventional XAI, which explains predictions at a given point, Delta-XAI frameworks provide directional, magnitude-sensitive, and context-aware explanations of prediction shifts, supporting robust, real-time decision-making in time-sensitive domains such as clinical monitoring, finance, and model lifecycle management (Kim et al., 3 Jul 2025, Kim et al., 28 Nov 2025, Carlo et al., 25 Jul 2024, Rida et al., 2023).

1. Conceptual Foundations and Scope

Delta-XAI encompasses any methodology that attributes prediction changes—either temporally (Δŷ between time steps), across model versions (Δmodel output), or conditional input shifts (Δfeature)—to specific features or rule sets, along with directionality (positive or negative impact) and context. The evolution from static XAI is motivated by domain needs where the absolute level of prediction is less actionable than its change (e.g., a drop in patient risk may warrant intervention) (Kim et al., 28 Nov 2025). Delta-XAI frameworks can be grouped as follows:

• Time-series prediction change explainers: DeltaSHAP, SWING, and generalized wrappers attributing changes between successive model outputs.
• Model comparison explainers: DeltaXplainer, which explicitly summarizes and rules where two classifiers disagree.
• Sensitivity-based explainers for local input changes: δ-XAI, which quantifies the impact of individual feature value changes on output likelihood.
• Decomposition-based explainers: DXAI, providing the distinction between class-agnostic and class-distinct parts in image data.

2. Delta-XAI Methods for Online Time Series: DeltaSHAP and SWING

DeltaSHAP is formulated to explain the causes underlying prediction evolutions in online patient monitoring. At each time $t$, the model output $f_t(x_{1:t})$ is explained not in isolation but via the change $\Delta f = f_{t+1}(x_{1:t+1}) - f_t(x_{1:t})$. The method adapts Shapley-value attribution to this setting, considering only actually observed feature combinations at $t+1$ to efficiently allocate $\Delta f$:

$$\phi_j(f, x_{1:t+1}) = \sum_{S \subseteq \mathfrak{F}\setminus\{j\}} \frac{|S|!(D-|S|-1)!}{D!} [v(S\cup\{j\}) - v(S)]$$

The sum of attributions is equal to the prediction change: $\sum_j \phi_j(f, x_{1:t+1}) = \Delta f$. In practice, feature attributions are computed via sampling and normalized to exactly match $\Delta f$, preserving both sign and scale (Kim et al., 3 Jul 2025).

The Shifted Window Integrated Gradients (SWING) algorithm extends gradient-based XAI to prediction changes, incorporating temporal dependencies and mitigating out-of-distribution baselines. SWING uses retrospective baseline selection and piecewise-linear historical integration to attribute $\Delta f$ with theoretical completeness and implementation invariance, outperforming adapted classical methods across extensive benchmarks (Kim et al., 28 Nov 2025).

Method	Core Principle	Attribution Target	Efficiency
DeltaSHAP	Shapley allocation of Δf	Time step prediction Δ	O(N·D)
SWING	Window-integrated gradients	Time-series Δŷ	~0.35s/sample

3. Model Comparison and Rule-Based Delta-XAI: DeltaXplainer

DeltaXplainer targets the model-selection lifecycle, attributing where and why two binary classifiers $f$ and $g$ differ. For a bounded feature space $\Omega\subseteq\mathbb{R}^d$, a "Δ-model" is learned to predict $\Delta^*(x) = 1$ iff $f(x)\neq g(x)$, and $0$ otherwise. This surrogate is realized as a structured set of axis-aligned rules $r(x)$ defining precise subspaces of disagreement (Rida et al., 2023). Each rule interprets the difference space, scored by coverage, precision, recall, and F1 metrics on held-out data. The approach yields human-readable, compact rule sets offering global transparency into model-drift scenarios.

4. Sensitivity-Based Δ-XAI: δ-XAI Index

The δ-XAI index extends global sensitivity analysis (GSA) to local instance-level explanations. Given a trained model $h:\mathbb{R}^M\rightarrow\mathbb{R}$, the per-feature score for instance $x^*$ with prediction $y^*=h(x^*)$ is:

$\delta_i = f_{Y}(y^*|X_i=x^*_i) - f_{Y}(y^*)$

Here, $f_{Y}(y^*|X_i=x^*_i)$ is the density of the output $y$ at $y^*$, conditional on feature $i$ fixed at $x^*_i$. Normalization yields a relative importance:

$I_i = |\delta_i| / \sum_{j=1}^M |\delta_j|$

δ-XAI highlights extreme or rare feature values and their impact on output likelihood, offering distribution-aware, moment-independent explanations robust to feature correlation. Computational complexity is $O(L·M)$ (bootstrap rounds × features) (Carlo et al., 25 Jul 2024).

5. Decomposition-Based Δ-XAI: DXAI for Images

The DXAI framework decomposes an input image $x$ into class-agnostic $\psi_{Agn}(x)$ and class-distinct $\psi_{Dist}(x)$ components:

$x = \psi_{Agn}(x) + \psi_{Dist}(x)$

The class-agnostic part is characterized by the classifier assigning a uniform output, while the distinct part contains all features relevant for class decision. DXAI uses branched style-transfer GANs to perform this analysis-synthesis decomposition, enforced by specialized loss functions balancing fidelity, distinctness, class relevance, and reconstruction (Kadar et al., 2023). The approach is most effective for image data with dense, additive class cues, such as color and texture, and yields robust, class-specific visualizations.

6. Evaluation Metrics, Empirical Results, and Theoretical Properties

Delta-XAI frameworks introduce specialized evaluation metrics to assess faithfulness, sufficiency, and temporal coherence of attributions:

• Cumulative Prediction Difference (CPD) and Cumulative Prediction Preservation (CPP) measure alignment between attributed feature removals and prediction changes.
• Area-based metrics (AUPD, AUPP, MPD/AUMPD, MPP/AUMPP) aggregate faithfulness across feature ranks and time windows.
• Pearson correlation (Corr) assesses alignment between features' ordering and true ∆predictions.
• Global class reduction/AUC evaluates decomposition methods by analyzing change in classification accuracy as class-distinct features are removed (Kadar et al., 2023).

Empirically, DeltaSHAP and SWING deliver substantial improvements in both explanation quality (up to 62% gain in faithfulness) and computational speed (33% time reduction) on large-scale clinical time-series datasets (Kim et al., 3 Jul 2025, Kim et al., 28 Nov 2025). δ-XAI demonstrates enhanced sensitivity to dominant and rare features compared to Shapley values, and DeltaXplainer reliably identifies compact disagreement rules in model drift scenarios (Carlo et al., 25 Jul 2024, Rida et al., 2023). DXAI achieves lower area-under-curve scores for class reduction and greater stability in pixel-wise decomposition compared to heatmap methods (Kadar et al., 2023).

7. Limitations, Extensions, and Applications

Delta-XAI methods are subject to domain-specific constraints. DeltaSHAP and SWING require time-series input structure; δ-XAI’s density estimation may be sensitive to sample size and bandwidth; DXAI incurs GAN training overhead and lacks direct pixel ranking. Rule-based methods (DeltaXplainer) are limited to numeric features and axis-aligned splits. Extensions under discussion include adaptive baseline selection, deeper integration with local and counterfactual explainers, genetic/global rule optimization, and exploration of non-linear or multi-class boundaries (Kim et al., 28 Nov 2025, Carlo et al., 25 Jul 2024, Rida et al., 2023).

Delta-XAI regimes are increasingly employed in clinical monitoring (real-time risk assessment), financial churn analysis, model drift diagnostics, and high-stakes automated decision systems. By providing actionable, context-grounded explanations of prediction changes, Delta-XAI enhances trustworthiness and interpretability, addressing critical gaps unaddressed by static XAI methodologies.