Corr-SHAP: Correlation-Aware Sampling for Faithful SHAP Value Estimation

In Human Activity Recognition (HAR), understanding the intricacy of body movements within high-risk applications is essential. This study uses SHapley Additive exPlanations (SHAP) to explain the decision-making process of Graph Convolution Networks (GCNs) when classifying activities with skeleton data. We employ SHAP to explain two real-world datasets: one for cerebral palsy (CP) classification and the widely used NTU RGB+D 60 action recognition dataset. To test the explanation, we introduce a novel perturbation approach that modifies the model's edge importance matrix, allowing us to evaluate the impact of specific body key points on prediction outcomes. To assess the fidelity of our explanations, we employ informed perturbation, targeting body key points identified as important by SHAP and comparing them against random perturbation as a control condition. This perturbation enables a judgment on whether the body key points are truly influential or non-influential based on the SHAP values. Results on both datasets show that body key points identified as important through SHAP have the largest influence on the accuracy, specificity, and sensitivity metrics. Our findings highlight that SHAP can provide granular insights into the input feature contribution to the prediction outcome of GCNs in HAR tasks. This demonstrates the potential for more interpretable and trustworthy models in high-stakes applications like healthcare or rehabilitation.

Brain tumors require multimodal therapies such as surgery, radiation treatment, and chemotherapy. There must be individual-level predictors that will influence treatment assignment patterns in the synthetic dataset to make precision medicine for neuro-oncology possible. The objective of this study is to demonstrate an interpretable machine learning (ML) approach for comparing predictors of surgery, radiation therapy, and chemotherapy in brain tumors using a synthetic dataset, with an emphasis on model comparison and SHapley Additive exPlanations (SHAP)-based explainability rather than clinical inference. The Kaggle dataset used in this study is fully synthetic and intended for methodological and educational research. Kaggle's 20,000 brain tumor patient data set was downloaded. Sixteen significant features were preprocessed and used for training nine supervised classifiers. The best performance was found using Gradient Boosting, and it was selected for SHAP analysis. SHAP beeswarm and waterfall plots were employed to interpret both global and local feature contributions. The Gradient Boosting model performed superior to all three tasks of treatment prediction—surgery, radiation therapy, and chemotherapy—when compared with five key classification metrics. SHAP analysis identified Tumor_Size, Tumor_Growth_Rate, and Age as the strongest predictors for treatments. For surgery, SHAP waterfall plots indicated that Tumor_Growth_Rate = − 1.694 had the highest negative impact (− 0.15), while Age = 0.612 and Symptom_2 = − 1.347 had positive impacts (+ 0.02 each). For radiation treatment, Location = 1.324 reduced treatment probability (− 0.05), while Tumor_Size had positive impact (+ 0.03) but Tumor_Growth_Rate = − 1.694 had a negative impact (− 0.03). Age = 0.612 was the most significant negative impact (− 0.04) in chemotherapy, with Tumor_Size = − 0.381 (− 0.03) and Family_History = − 1 following closely behind. Tumor_Growth_Rate contributed a small level of enhanced prediction (+ 0.01). These findings demonstrate the feasibility of using interpretable machine learning methods to explore treatment-related patterns in healthcare-like data and provide a foundation for future validation using real-world clinical datasets.

Explainable AI (XAI) is essential for fostering trust in the predictions of deep neural networks (DNNs), especially within the domain of image classification. SHAP (SHapley Additive exPlanations), a theoretically sound method rooted in game theory, stands as a prominent XAI technique for attributing feature importance. However, SHAP’s inherent computational complexity, which grows exponentially with the number of features, restricts its practical application. Although numerous approximation methods have been proposed, they often deviate from the original SHAP formulation, potentially sacrificing accuracy and theoretical fidelity. In this paper, we introduce CorrSHAP, a novel approach that leverages image superpixel correlations to significantly accelerate SHAP value estimation while preserving the rigor of the original formulation. CorrSHAP efficiently quantifies the interdependence of superpixels within an image, enabling the SHAP calculation to consider only interdependent superpixel combinations. This targeted approach significantly reduces the computational burden without compromising the faithfulness of the explanation. Evaluation results demonstrate that CorrSHAP outperforms the comparable Monte Carlo SHAP method in explanation faithfulness, achieving this in a fraction of a second, with a 55-fold speed improvement. The source code is available on the link (https://github.com/vhasic/CorrSHAP).

Explaining machine learning (ML) models using eXplainable AI (XAI) techniques has become essential to make them more transparent and trustworthy. This is especially important in high-risk environments like healthcare, where understanding model decisions is critical to ensure ethical, sound, and trustworthy outcome predictions. However, users are often confused about which explanability method to choose for their specific use case. We present a comparative analysis of two explainability methods, Shapley Additive Explanations (SHAP) and Gradient-weighted Class Activation Mapping (Grad-CAM), within the domain of human activity recognition (HAR) utilizing graph convolutional networks (GCNs). By evaluating these methods on skeleton-based input representation from two real-world datasets, including a healthcare-critical cerebral palsy (CP) case, this study provides vital insights into both approaches’ strengths, limitations, and differences, offering a roadmap for selecting the most appropriate explanation method based on specific models and applications. We qualitatively and quantitatively compare the two methods, focusing on feature importance ranking and model sensitivity through perturbation experiments. While SHAP provides detailed input feature attribution, Grad-CAM delivers faster, spatially oriented explanations, making both methods complementary depending on the application’s requirements. Given the importance of XAI in enhancing trust and transparency in ML models, particularly in sensitive environments like healthcare, our research demonstrates how SHAP and Grad-CAM could complement each other to provide model explanations.

IntroductionThe need for eXplainable Artificial Intelligence (XAI) in healthcare is more critical than ever, especially as regulatory frameworks such as the European Union Artificial Intelligence (EU AI) Act mandate transparency in clinical decision support systems. Post hoc XAI techniques such as Local Interpretable Model-Agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDPs) are widely used to interpret Machine Learning (ML) models for disease risk prediction, particularly in tabular Electronic Health Record (EHR) data. However, their reliability under real-world scenarios is not fully understood. Class imbalance is a common challenge in many real-world datasets, but it is rarely accounted for when evaluating the reliability and consistency of XAI techniques.MethodsIn this study, we design a comparative evaluation framework to assess the impact of class imbalance on the consistency of model explanations generated by LIME, SHAP, and PDPs. Using UK primary care data from the Clinical Practice Research Datalink (CPRD), we train three ML models: XGBoost (XGB), Random Forest (RF), and Multi-layer Perceptron (MLP), to predict lung cancer risk and evaluate how interpretability is affected under class imbalance when compared against a balanced dataset. To our knowledge, this is the first study to evaluate explanation consistency under class imbalance across multiple models and interpretation methods using real-world clinical data.ResultsOur main finding is that class imbalance in the training data can significantly affect the reliability and consistency of LIME and SHAP explanations when evaluated against models trained on balanced data. To explain these empirical findings, we also present a theoretical analysis of LIME and SHAP to understand why explanations change under different class distributions. It is also found that PDPs exhibit noticeable variation between models trained on imbalanced and balanced datasets with respect to clinically relevant features for predicting lung cancer risk.DiscussionThese findings highlight a critical vulnerability in current XAI techniques, i.e., their interpretability are significantly affected under skewed class distributions, which is common in medical data and emphasises the importance of consistent model explanations for trustworthy ML deployment in healthcare.

Integrating multi-source remote sensing can improve the accuracy of forest aboveground biomass (AGB) estimation. However, the accuracy and stability of the forest AGB estimation results are affected by multiple remote sensing feature variables as well as parameter tuning of machine learning algorithms. To this end, this study employed six types of remote sensing data—Landsat 8 OLI, Sentinel-2A, GEDI, ICESat-2, ALOS-2, and SAOCOM. A dual-variable selection strategy based on SHapley Additive exPlanations (SHAP) was developed, and a genetic algorithm (GA) was used to optimize the parameters of five machine learning models—elastic net (EN), least absolute shrinkage and selection operator (Lasso), support vector regression (SVR), Random Forest (RF), and Categorical Boosting (CatBoost)—to estimate the AGB of Pinus kesiya var. langbianensis forest in Wuyi Village, Zhenyuan County. The dual-variable selection strategy integrates SHAP with the Pearson correlation coefficient (PC), RF, EN, and Lasso to enhance feature screening robustness and interpretability. The results of the study showed that Lasso-SHAP dual-variate screening was more stable than SHAP univariate screening. In particular, the Lasso-SHAP strategy improved the average R2 from 0.59 (using SHAP alone) to above 0.70, achieving an enhancement of 11%. Among GA-optimized parametric machine learning models, the linear GA-Lasso achieved the best performance, with an R2 of 0.91 and an RMSE of 12.94 Mg/ha, followed by the GA-EN model (R2 = 0.89, RMSE = 14.46 Mg/ha). For nonlinear models, GA-SVR performed the best (R2 = 0.74, RMSE = 22.07 Mg/ha), surpassing the GA-CatBoost model (R2 = 0.64, RMSE = 25.88 Mg/ha). In summary, the Lasso-SHAP dual-variable selection strategy effectively improves the estimation accuracy of AGB for Pinus kesiya var. langbianensis forests, while GA-optimized machine learning models demonstrate excellent performance, providing strong support for regional-scale forest resource monitoring and carbon stock assessment.

Electricity theft remains a significant challenge for modern power systems causing major financial damage to utility providers. With increasing adoption of advanced metering infrastructure in smart grids, energy efficiency and real-time monitoring have improved but the vulnerability of smart meters to tampering and cyber manipulation introduces new threats. To address these concerns at scale, this study proposes a cloud-based machine learning framework for intelligent electricity theft detection in residential sectors. The cloud-based setup, utilizing a Google Colab environment, facilitates centralized data processing, streamlined deployment, and real-time decision support. This study systematically resolves several critical limitations in existing approaches. Firstly, it addresses the severe class imbalance in real-world datasets, which significantly underrepresent theft cases compared to honest consumption. The synthetic minority oversampling technique is used to equalize representation and enable fair learning. Second, recognizing that synthetic oversampling can introduce noise and class overlap, a post-balancing noise reduction mechanism is integrated to refine the dataset and eliminate ambiguous samples. Third, by combining the interpretability of Decision Trees (DT) and the robustness of Random Forests (RF), an ensemble model named DTRF is proposed to enhance variance control and generalization. To improve the predictive performance of DTRF, the Firefly Algorithm (FA) is used to smartly adjust the model’s settings based on the specific features of the dataset. The proposed models are tested on two publicly available real-world datasets, Pakistan Residential Electricity Consumption (PRECON) and State Grid Corporation of China (SGCC), showing consistent improvements over standard models in various measures. On the PRECON dataset, the DTRF model outperforms baseline models and achieves a 25.68% reduction in log loss, a 31.23% decrease in Hamming loss, a 3.52% improvement in Cohen’s Kappa (CK), and a 4.10% increase in Jaccard score. The proposed DTRF-FA model further achieves a 40.97% reduction in log loss, a 36.38% decrease in Hamming loss, a 2.71% improvement in CK, and a 3.17% increase in the Jaccard score relative to the DTRF model. On the SGCC dataset, the DTRF model outperforms baselines with a 33.34% drop in Hamming loss, a 3.86% increase in CK, a 3.70% rise in the Jaccard score, and a 1.92% gain in the F1-score. The DTRF-FA model achieves a 75.00% reduction in Hamming loss, a 5.63% gain in CK, a 5.82% rise in Jaccard score, a 2.95% improvement in F1-score, and a 5.39% increase in Matthews correlation coefficient over the DTRF baseline. These improvements affirm the models’ ability to balance sensitivity and specificity while maintaining low false positive and false negative rates. Finally, to promote interpretability and transparency essential for stakeholder trust and regulatory compliance, the framework integrates two powerful explainable artificial intelligence techniques. Local Interpretable Model-agnostic Explanations (LIME) is used to generate local, instance-level explanations of individual predictions, while SHapley Additive exPlanations (SHAP) provides a global view of feature importance using cooperative game theory. By combining LIME and SHAP, the framework helps practitioners understand models’ decisions in detail and as a whole ensuring that it is accurate, clear, adaptable, and suitable for use in the smart grid settings.

A detailed analysis of temperature and relative humidity biases in ERA5 reanalysis under tropical cyclone (TC) conditions is conducted using a composite analysis method based on more than 17,000 dropsonde observations. The statistical results indicate that ERA5 underestimates the warm‐core intensity of TCs. It shows a moist bias at the tropopause and a cold‐moist bias in the boundary layer beyond four times the radius of maximum winds (RMW). In contrast, within 4 RMW, a warm‐dry bias in the boundary layer was observed. The temperature bias is more pronounced at the TC center, and relative humidity exhibits a similar pattern, primarily in the mid and upper troposphere. The bias distribution is asymmetric, most prominent in the right front quadrant of the TC's motion direction. Besides, bias correction is performed using an eXtreme Gradient Boosting (XGBoost) model and the biases can be reduced by approximately 60%. The dependence relationships between relative humidity and the influencing factors are investigated through SHapley Additive exPlanations (SHAP) values. The results indicate that low‐level relative humidity increases and instability intensifies in TCs at night. Additionally, the findings suggest that ERA5 may inadequately capture the diurnal variation of relative humidity in TCs. The SHAP analysis also shows that relative humidity tends to be stronger in TC's motion direction.

Banking and online financial service providers face significant challenges due to financial fraud. Traditional frauddetection methods are often inadequate because of imbalanced datasets, limited interpretability, and privacy concerns involving confidential customer information. This paper presents an explainable AI–based system for financial fraud detection designed to address these issues. The system employs the Light Gradient Boosting Machine (LightGBM) as the primary model, combined with SMOTE oversampling to mitigate class imbalance. Privacy is maintained by anonymizing sensitive features, including Personally Identifiable Information (PII), by temporarily adding and later removing attributes such as name_email_similarity before model training. Model transparency is achieved through SHAP (Shapley Additive Explanations), which offers featurelevel interpretability for fraud predictions. The system is implemented as a web-based interactive dashboard using the Flask framework, enabling users to upload datasets, perform fraud detection, adjust detection sensitivity (via threshold tuning), and download a detailed fraud report. When evaluated on a real-world dataset, the system achieved an overall accuracy of 98.5%, an ROC-AUC of 0.89, improved privacy preservation, and enhanced interpretability through SHAP. The proposed solution provides a practical end-to-end framework that balances accuracy, transparency, and privacy protection, making it suitable for banking and fintech fraud-detection applications.

Developing fuel resources is strategically crucial for Armenia. Far more than any other fossil fuel resource, coal roughly generates half the nation’s electricity. Although coal could play a critical role, no vast data is available about Armenia coal properties. Using robust modeling of energy indexes such as coal gross calorific value (GCV) by considering trivial existing datasets could be an essential clue for ensuring sustainable development. For the first time, this investigation is going to model GCV for Armenia coal samples. For this purpose, SHAP (SHapley Additive exPlanations) as a novel explainable artificial intelligence will be introduced. SHAP enables understanding the magnitude of relationships between each individual input record and its representative output and ranks input variables based on their effectiveness. SHAP was coupled by extreme gradient boosting (xgboost) as the most recently generated powerful predictive machine learning tool (SHAP-Xgboost). SHAP-Xgboost could accurately (R2=0.99) model GCV based on proximate and ultimate variables of Armenia coal samples. These significant outcomes open a new window for developing high interpretability models to assess coal properties and pinpoint the influential parameters.

This paper presents a two-stage explainable recommendation system for reciprocal affiliate–seller matching that uses machine learning and data science to handle voluminous data and generate personalized ranking lists for each user. In the first stage, a representation learning model was trained to create dense embeddings for affiliates and sellers, ensuring efficient identification of relevant pairs. In the second stage, a learning-to-rank approach was applied to refine the recommendation list based on user suitability and relevance. Diversity-enhancing reranking (maximal marginal relevance/explicit query aspect diversification) and popularity penalties were also implemented, and their effects on accuracy and provider-side diversity were quantified. Model interpretability techniques were used to identify which features affect a recommendation. The system was evaluated on a fully synthetic dataset that mimics the high-level statistics generated by affiliate platforms, and the results were compared against classical baselines (ALS, Bayesian personalized ranking) and ablated variants of the proposed model. While the reported ranking metrics (e.g., normalized discounted cumulative gain at 10 (NDCG@10)) are close to 1.0 under controlled conditions, potential overfitting, synthetic data limitations, and the need for further validation on real-world datasets are addressed. Attributions based on Shapley additive explanations were computed offline for the ranking model and excluded from the online latency budget, which was dominated by approximate nearest neighbors-based retrieval and listwise ranking. Our work demonstrates that high top-K accuracy, diversity-aware reranking, and post hoc explainability can be integrated within a single recommendation pipeline. While initially validated under synthetic evaluation, the pipeline was further assessed on a public real-world behavioral dataset, highlighting deployment challenges in affiliate–seller platforms and revealing practical constraints related to incomplete metadata.

Accurate house price prediction is crucial for stakeholders in the real estate industry, including buyers, sellers, agents, and investors. Traditional valuation methods often rely on manual appraisal or basic regression techniques, which may lack scalability and the ability to capture complex relationships among features. With the rise of machine learning, especially ensemble models like XGBoost, there is a growing opportunity to improve prediction accuracy and robustness. This study presents a comprehensive framework for optimal house price prediction using XG-Boost, applied to the King County housing dataset comprising over 21,000 records. The suggested objective is supported by carefully conducted data preprocessing, which will imply outliers’ elimination, feature engineering, and normalization. Other important derived features that were included were the age of the house and the renovation status of the house as well as the price per square foot. The XG-Boost regression model was trained and validated, and the R-squared ratio was found to be 0.87, indicating decent model performance. The model was also deployed as an interactive web application with Streamlit to allow users to store property details and get real-time predictions of the properties. Also, Shapley Additive Explanations (SHAP) values were applied to interpret the model output in a way that enhances explanations and builds trust among users. The system is not only accurate, but also keen on accountability, and usability. The proposed research will fill this missing gap and provide an actionable machine learning solution, deployable and explainable to real estate businesses to take advantage of the predictive analytics revolution. The results demonstrate the effectiveness of ensemble learning in conjunction with visualization and interpretability tools in the development of robust decision-support systems in the real-world housing markets.

Cost estimation is a key component of project plans, yet it is challenging to provide reliable and efficient estimations using conventional methods in the conceptual phase of infrastructure projects. This study proposes a framework that integrates feature selection, extreme gradient boosting (XGBoost), Bayesian optimization (BO), and SHapley Additive exPlanations (SHAP) to provide conceptual cost estimations and explain the results for early decision-making. Correlation analysis and forward search are combined to select the key features. XGBoost is developed as the estimator and enhanced by BO in accuracy and efficiency. Model explanations were presented using SHAP. The framework is demonstrated through a case study of electric substations containing 605 samples. The results show that the proposed framework can provide satisfactory performance on conceptual cost estimations, where BO-XGBoost outperforms the benchmark models (with R 2 ~0.9567, adjusted R 2 ~0.9549, RMSE ~ 0.8690, and MAE ~ 0.4875). SHAP reveals how the features contribute to the cost based on both global and local explanations. The framework provides a guideline for more accurate, efficient, and explainable cost estimations in the conceptual phase of infrastructure projects. It can support the government and project planners in early decision-making, including reliable project budget and plan alternatives selection.

Precision medicine relies on accurate and interpretable predictive models to identify patient subgroups and biomarkers that can guide individualized treatment strategies. While extreme gradient boosting (XGBoost) often achieves state-of-the-art predictive performance, its complexity can impede understanding of how input variables influence outcomes. Building upon existing XGBoost frameworks for estimating individualized treatment rule (ITR), we introduce a global permutation test within this framework to assess treatment effect heterogeneity. Additionally, we incorporate two model-agnostic explanation techniques, local interpretable model-agnostic explanations (LIME) and SHapley Additive exPlanations (SHAP), to enhance interpretability at both global and individual levels. Through simulations and analyses of real-world clinical trial datasets, we illustrate that our permutation-based pipeline can detect empirical signals of treatment effect heterogeneity, while LIME and SHAP offer exploratory insights into feature contributions and ITR.

Federated learning with explainable AI for liver disease prediction: A privacy-preserving approach