Correlation-Aware Feature Attribution Based Explainable AI (2511.16482v1)

Abstract: Explainable AI (XAI) is increasingly essential as modern models become more complex and high-stakes applications demand transparency, trust, and regulatory compliance. Existing global attribution methods often incur high computational costs, lack stability under correlated inputs, and fail to scale efficiently to large or heterogeneous datasets. We address these gaps with \emph{ExCIR} (Explainability through Correlation Impact Ratio), a correlation-aware attribution score equipped with a lightweight transfer protocol that reproduces full-model rankings using only a fraction of the data. ExCIR quantifies sign-aligned co-movement between features and model outputs after \emph{robust centering} (subtracting a robust location estimate, e.g., median or mid-mean, from features and outputs). We further introduce \textsc{BlockCIR}, a \emph{groupwise} extension of ExCIR that scores \emph{sets} of correlated features as a single unit. By aggregating the same signed-co-movement numerators and magnitudes over predefined or data-driven groups, \textsc{BlockCIR} mitigates double-counting in collinear clusters (e.g., synonyms or duplicated sensors) and yields smoother, more stable rankings when strong dependencies are present. Across diverse text, tabular, signal, and image datasets, ExCIR shows trustworthy agreement with established global baselines and the full model, delivers consistent top-$k$ rankings across settings, and reduces runtime via lightweight evaluation on a subset of rows. Overall, ExCIR provides \emph{computationally efficient}, \emph{consistent}, and \emph{scalable} explainability for real-world deployment.

• The paper introduces ExCIR, a correlation-aware method that computes the Correlation Impact Ratio (CIR) for efficient and consistent feature attribution.
• It achieves near-perfect feature ranking with only 20–40% of data, delivering speedups up to 100× compared to SHAP and LIME.
• The method integrates BlockCIR to aggregate correlated features, stabilizing rankings and supporting streaming, low-memory applications.

ExCIR: Correlation-Aware Feature Attribution for Explainable AI

Introduction and Motivation

Feature attribution is foundational for Explainable AI (XAI), powering transparency, trust, auditing, and compliance for ML models deployed in regulated or safety-critical domains. Conventional global attribution frameworks, including perturbation-based (LIME, SHAP, PFI, PDP-variance) and gradient-based approaches (saliency maps, Integrated Gradients), often suffer from substantial computational overhead, lack resilience to cross-feature dependencies, and may double-count in the presence of collinear or redundant features. These limitations can impair interpretability and scalability, especially for large, heterogeneous datasets or streaming applications. The "Correlation-Aware Feature Attribution Based Explainable AI" (ExCIR) (2511.16482) addresses these challenges via a robust, computationally efficient, correlation-aware attribution schema, providing consistent and scalable global feature rankings without repeated model evaluation.

ExCIR Methodology

ExCIR computes the Correlation Impact Ratio (CIR), quantifying the sign-aligned co-movement between feature sets and model outputs after robust centering. The central pipeline consists of robust centering (via mid-mean or median), single-pass accumulation of co-movement primitives, and blockwise aggregation for correlated features (BlockCIR). CIR delivers translation and scale invariance, sign symmetry, and monotonicity, operating in $O(nd)$ time. BlockCIR aggregates related features (user-specified or data-driven), effectively mitigating signal redundancy and stabilizing rankings in collinear groups.

Figure 1: ExCIR heat-map agreement—rank overlap, Procrustes shape alignment, and symmetric KL—showing lightweight (20–40% rows) runs closely preserve full-data rankings and score distributions.

Class-conditioned ExCIR integrates per-class scoring for multiclass models, utilizing logits or calibrated scores, and robustly centering the features and outputs before aggregation. The protocol is inherently streaming-friendly, leveraging quantile sketches for low-memory scenarios.

Lightweight Transfer Protocol

To further enhance computational practicality, ExCIR includes a lightweight transfer mechanism: global feature rankings can be reliably reconstructed with only 20–40% of the dataset, yielding top-$k$ Jaccard overlap and low shape/distributional residuals versus full-data runs. Even modest subsampling stabilizes CIR vectors due to reliance on second-order signal structure—initial samples capture the essential correlation patterns which are subsequently refined.

Figure 2: Runtime scaling across datasets illustrating near-linear runtime reduction with subsample fraction $f$; robust speed-ups, especially for signal/image tasks.

Figure 3: Runtime versus Jaccard@8 Pareto frontier; agreement saturates for $f$ in $[0.2, 0.4]$, enabling large time savings.

Empirical Evaluation and Diagnostics

ExCIR was systematically validated on 29 benchmarks spanning text, image, tabular, signal, and synthetic modalities using fixed model architectures (e.g., logistic regression, XGBoost). Across all domains, ExCIR reproduces full-model rankings with high fidelity using only a fraction of rows. For instance, with 20% of data, ExCIR achieves near-perfect rank overlap (Jaccard@k) and low Procrustes residuals, demonstrating strong cost–agreement trade-offs. Benchmarking against SHAP and LIME, ExCIR yields Kendall-$\tau$ agreements of $0.82$ and $0.51$, respectively, and is approximately $40\times$ faster than LIME, and $100\times$ faster than SHAP.

Figure 4: Kendall–$\tau$ agreement for subsampled ExCIR vs. full-data reference; smooth convergence to $\tau=1.0$ at full fraction.

Feature-importance distributions confirm that ExCIR produces compact, reproducible profiles across modalities, aligning well with MI, gain-based, and surrogate-LR methods. As a correlation-based metric, ExCIR is robust to redundant features; groupwise BlockCIR further reduces double-counting, preserving top-$k$ features.

Figure 5: Beeswarm visualization of per-feature scores for a representative tabular dataset.

Ablation analyses show centering choice (mid-mean, median) does not substantially alter rankings; ExCIR remains deterministic and scalable.

Figure 6: Centering ablation—Spearman correlation with mid-mean centering and runtime ratios; mid-mean/median remain robust.

Figure 7: Groupwise aggregation—BlockCIR profiles and top-$k$ Jaccard overlap on correlated features, demonstrating mitigation of double-counting.

Limitations and Theoretical Implications

Diagnostic and ground-truth recovery experiments indicate that ExCIR scores meaningful attributions, but, as a correlation-based method, it should not be interpreted as a causal explainer. In the presence of strong causal confounding or nonlinear dependence, CIR can over-credit correlated, non-causal features. Whitening and reporting BlockCIR groups address multicollinearity, but users must interpret CIR in the context of global association, not causality.

Numerical highlights include:

• Achieving Jaccard@8 overlap $\geq 0.9$ with only 20–40% of the data,
• CPU speed-ups of $3$–$9\times$ across tasks,
• Strong agreement with SHAP and surrogate gain/MI methods ($\tau_{SHAP}=0.82$).

Practical Implications and Future Directions

ExCIR's lightweight, streaming-friendly construction makes it suitable for deployment in latency-constrained, privacy-sensitive, or edge environments. The architecture-agnostic, translation- and scale-invariant design ensures consistent global attribution reporting across retraining, calibration shifts, and heterogeneous data sources. For domains with correlated features, BlockCIR enables stable, interpretable groupwise rankings.

The theoretical foundations (translation/scale invariance, sign symmetry, monotonicity) underpin its resilience, while the closed-form computation facilitates adoption in real-time or large-scale workflows. The method's generalization to class-conditioned, weighted, and groupwise settings is fully formalized, supporting extension to emerging data modalities.

Future research will pursue Conditional ExCIR (cCIR) for disentangling unique contributions in correlated blocks and Mutual-ExCIR (mCIR) for nonlinear dependence capture via MI frameworks, maintaining computational efficiency.

Conclusion

ExCIR provides a formal, correlation-aware, and computationally efficient feature attribution framework. Diagnostic analyses and large-scale benchmarking demonstrate that ExCIR stably reproduces global model rankings—with minimal data and compute—across diverse domains. BlockCIR broadens coverage to correlated features, reducing redundant signal attribution. While limited as a causal indicator, ExCIR robustly serves real-world explainability needs, complementing model-faithful gain/kernel methods and enabling interpretable, reproducible, and scalable deployments.