shap analysis

In multi-omics deep learning, feature attribution results exhibit substantial sensitivity to model architecture and random weight initialization, undermining the consistency and robustness of critical biomolecule identification. This work systematically evaluates the stability of attribution methods—particularly SHAP—across multi-view deep learning models, revealing significant rank-order divergence across architectures and initialization seeds. To address this, we propose a robustness diagnostic framework that replaces single attribution outputs with subset-based modeling and clustering-quality assessment, quantifying the consistency of feature importance estimates. Experiments on benchmark multi-omics datasets demonstrate that our approach reliably identifies high-stability biomarker candidates, markedly improving the reproducibility and trustworthiness of interpretability analyses. By providing a principled, quantifiable measure of attribution robustness, the framework establishes a more reliable foundation for model-driven biological discovery.

This work addresses the longstanding challenge of jointly optimizing model interpretability and predictive performance in machine learning. We propose a SHAP-driven regularized training framework, whose core innovation is the first direct incorporation of TreeSHAP attribution values into the loss function via a novel joint regularization term based on the entropy of the attribution distribution. This term simultaneously enforces sparsity, concentration, and cross-sample stability of feature importances. Unlike post-hoc methods, our approach is end-to-end trainable, applicable to mainstream tree-based models (e.g., XGBoost, LightGBM), and supports both regression and classification tasks. Extensive experiments across multiple benchmark datasets demonstrate that the proposed method improves model generalization, yields more robust and interpretable SHAP attributions, and maintains or exceeds baseline accuracy—without sacrificing predictive performance.

SHAP struggles to distinguish causal from spurious correlations, leading to erroneous feature attributions under high feature collinearity. To address this, we propose Causal-SHAP—a novel framework that integrates the PC algorithm (for causal graph discovery) and the IDA algorithm (for interventional causal effect estimation) into the SHAP pipeline, enabling intervention-based correction of feature importance scores. By explicitly modeling the underlying causal structure among variables, Causal-SHAP preserves SHAP’s local interpretability while suppressing attribution to non-causal (i.e., spuriously correlated) features. Extensive experiments on synthetic data and multiple real-world benchmark datasets demonstrate that Causal-SHAP significantly reduces importance scores for non-causal features—by an average of 37.2%—thereby enhancing the causal fidelity of attributions and strengthening the practical utility of model explanations for decision support.

Existing feature attribution methods (e.g., SHAP) in information retrieval provide only document-level pointwise explanations, failing to capture inter-document relative ranking relationships within a ranked list. Method: This paper formally defines the list-level feature attribution problem and proposes a Shapley-value-based theoretical framework for joint attribution over entire ranking outputs. We introduce two novel evaluation paradigms to assess attribution correctness and completeness; identify contrastive decision-making as a fundamental constraint on attribution design; develop LTR-model-adapted attribution algorithms; and propose explanation-driven qualitative validation techniques. Results: Experiments on standard LTR benchmarks demonstrate that our method precisely identifies features governing relative document positioning, overcoming inherent limitations of selection-based explanations and significantly enhancing the interpretability of ranking models.

This study addresses the troubling inconsistency in feature attribution methods such as SHAP, which can yield substantially divergent explanations even for identical inputs and models, thereby undermining trustworthiness and auditability in high-stakes applications. The work formally defines and quantifies the phenomenon of “explanation multiplicity,” distinguishing its origins in model training or selection from inherent randomness in the explanation procedure itself. To assess stability, the authors introduce a dual-perspective metric incorporating both feature magnitude and ranking, establish a randomized null model as an interpretable baseline, and develop a comprehensive empirical evaluation framework spanning diverse datasets and model classes. Experiments demonstrate that explanation multiplicity is pervasive; relying solely on SHAP value magnitudes can lead to misleading conclusions, necessitating rank-sensitive metrics and principled baselines for reliable interpretability assessment.

This study addresses the limited clinical adoption of machine learning models for Alzheimer’s disease (AD), which stems from the lack of robustness and consistency in existing interpretability methods across tasks, disease stages, and model architectures. To overcome this, we propose the first multidimensional SHAP evaluation framework that integrates coherence, stability, and cross-task consistency, systematically validating its transferability in both AD diagnosis and prognosis tasks. Leveraging AutoML-optimized classifiers combined with SHAP analysis, we quantitatively assess feature importance through feature correlation, top-k overlap ratio, sign consistency, and domain contribution ratios. Results demonstrate that cognitive and functional biomarkers predominantly drive explanations in both tasks, with SHAP attributions showing high consistency between diagnostic and prognostic models—achieving 100% sign stability and stable domain contributions, except for a modest increase in the contribution of genetic features in prognosis.

This study addresses the strong model dependency of SHAP value interpretations, which lacks a standardized analytical framework and thereby limits reliable explanations of black-box model decisions in high-stakes applications. For the first time, this work systematically evaluates SHAP explanations across multiple mainstream machine learning models on diverse datasets, uncovering consistent patterns of model dependence. Furthermore, it proposes a generalized waterfall plot visualization method tailored for multi-class classification problems. Experimental results demonstrate the effectiveness and practical utility of the proposed approach, offering both theoretical grounding and actionable guidance for practitioners in the field of explainable artificial intelligence.

This study addresses the challenge of poor response to conventional therapies in non-small cell lung cancer (NSCLC) due to tumor heterogeneity by proposing an interpretable predictive framework that integrates multi-omics data with machine learning. Leveraging patient-specific molecular profiles, the approach employs an XGBoost regression model optimized via random search and cross-validation to accurately predict drug sensitivity (LN-IC50). Key predictive features are identified using SHAP values, and biological interpretability is enhanced through semantic explanations generated by the large language model DeepSeek. The model demonstrates superior predictive performance while successfully uncovering genes and pathways associated with drug response, thereby significantly improving both interpretability and clinical credibility of the results.

Existing local attribution methods, such as Local Shapley Values and LIME, remain prone to erroneously assigning importance to irrelevant features even under ideal conditions, leading to unreliable explanations. This work proposes R-LOCO, a novel approach that partitions the input space into regions where feature importance is approximately homogeneous and applies a global attribution strategy within each region to construct instance-specific yet stable local explanations. By innovatively integrating local and global explanation paradigms, R-LOCO satisfies established axiomatic principles of sound attribution. Empirical evaluations demonstrate that R-LOCO significantly outperforms Local Shapley Values and LIME in both fidelity and stability, offering more trustworthy and consistent interpretability for individual predictions.

This work addresses the instability of feature attribution methods—often caused by variations in training splits and random seeds—which undermines model interpretability and decision reliability. To this end, the authors propose RoSHAP, a robust attribution framework that explicitly models attribution randomness through statistical distribution estimation. RoSHAP leverages bootstrap resampling and kernel density estimation to characterize the distribution of SHAP values and introduces an integrated metric that jointly quantifies feature activity, strength, and stability. Theoretical analysis establishes its asymptotic Gaussianity while substantially reducing computational overhead. Empirical results demonstrate that RoSHAP outperforms single-run attribution approaches in identifying true signal features, and that feature subsets selected via RoSHAP maintain predictive performance comparable to full-feature models despite using fewer variables.