ABSTRACT Healthcare sustainability assessment is a challenging multi‐criteria decision‐making (MCDM) problem due to uncertainty, heterogeneous data distributions and complex interactions among evaluation criteria. To solve these problems, a novel Granular Density‐based Relative β ‐Covering Approximation Space (GDR β CAS) is proposed and an integrated GDR β CAS‐VIKOR framework is constructed for decision analysis under uncertainty in this paper. The proposed method is an extension of the classical β ‐covering rough sets, which considers the local distribution properties in the approximation process. Specifically, relative β ‐covering neighbourhoods are constructed to capture adaptive granular relationships among alternatives, while novel granular density and relative granular distance measures are introduced to quantify local structural information. Furthermore, a reliability factor that integrates approximation accuracy and granular density is developed to simultaneously characterize uncertainty and distribution‐aware knowledge. This mechanism enables the approximation process to adapt to heterogeneous granular environments and provides a richer representation of alternatives than existing covering‐based rough set models. The resulting reliability information is incorporated into the VIKOR method, enabling the framework to consider not only criterion performance but also local uncertainty distribution and structural reliability, which improves discrimination among closely competing alternatives. The effectiveness of the proposed framework is demonstrated through a healthcare sustainability assessment involving 16 European countries. Comparative analyses with TOPSIS, VIKOR, COPRAS, MABAC and CODAS confirm the consistency and discriminative capability of the proposed approach, while sensitivity analyses with respect to , , and verify its robustness and stability. The results demonstrate that the GDR β CAS–VIKOR framework provides an effective and reliable decision‐support tool for complex MCDM problems under uncertainty.
ABSTRACT Depressed individuals with co‐occurring radical tendencies pose greater potential risk than general depression sufferers. Previous studies largely detect depression sufferers or classify depressive symptoms in social networks, ignoring the radicalisation process under long‐term complex emotions. Thus, we propose TFE‐AOD, a property‐specific method that incorporates temporal fine‐grained emotions into aggressive outbursts to enhance detection of depression individuals with radical tendencies. The proposed method first classifies fine‐grained emotions via an attention‐based Transformer, introducing irony (e.g., self‐mockery, sarcasm, struggle and powerlessness) to capture explicit depressive traits. It then combines fine‐grained emotions with nocturnal activity to identify depression in long text sequences, leveraging time‐dependent localisation to align with symptom accumulation. Nocturnal activity reflects the higher nighttime complaint propensity of depressed individuals. Finally, a knowledge‐enhanced pretrained model detects co‐occurrence by incorporating aggressive outburst features (alcohol abuse, drug dependency, covert hateful mentality, violence exposure). Experiments on Sina Weibo (SWDD) and a self‐constructed Douban dataset show that TFE‐AOD outperforms state‐of‐the‐art models in detecting this co‐occurrence.
ABSTRACT Accurate identification of signet ring cells (SRCs) in histopathological images is essential for the auxiliary evaluation of signet ring cell carcinoma (SRCC). Although deep learning has greatly advanced automated medical image interpretation, reliable SRC segmentation remains difficult because pixel‐level annotations are limited and SRCs often appear in crowded, adherent, or overlapping patterns. These characteristics increase the complexity of distinguishing individual cellular regions from surrounding tissues. To address these issues, this work proposes RGD‐UNet, a dedicated semantic segmentation network designed for SRC analysis. The model adopts an encoder–decoder framework and is optimized to capture discriminative pathological features while maintaining efficient computation. Unlike methods that rely on additional refinement procedures, RGD‐UNet can directly delineate clustered and overlapping SRC regions, thereby simplifying the segmentation workflow and improving practical usability. Furthermore, to facilitate the development and evaluation of SRC segmentation algorithms in clinical‐oriented scenarios, the DigestPath 2019 dataset was further extended by adding complete SRC mask annotations. Experimental results demonstrate that RGD‐UNet achieves superior segmentation precision compared with representative state‐of‐the‐art approaches while retaining low computational cost. These findings suggest that the proposed method has strong potential for assisting computer‐aided pathological diagnosis of SRCC.
ABSTRACT Current methods for multi‐trait essay scoring mainly focus on style‐based traits such as grammar and structure, with limited attention to multiple content‐level traits. This limitation prevents them from effectively assessing the fine‐grained content quality of essays, resulting in evaluation inaccuracies where high‐scoring essays may reflect formulaic writing rather than insightful content. In this paper, we propose a novel multi‐trait scoring task based on teachers' reflective essays, aiming to evaluate instructional insight conveyed through concrete content. To explore content‐oriented scoring of teaching reflection, we first construct a real‐world dataset, PTRE, consisting of 1820 reflective essays written by pre‐service teachers. Each essay is annotated with scores for four instructional content traits: Textual Expression, Problem Awareness, Comprehensive Thinking and Continuous Growth. Additionally, the dataset includes partial sentence‐level annotations highlighting trait‐relevant content as supporting evidence for each trait. We further design TMoLE, a topic‐driven mixture of experts framework for content‐oriented multi‐trait scoring. It learns trait‐specific topic distributions guided by seed words derived from sentence‐level annotations under expert guidance, ensuring semantic alignment with each scoring trait. Using these distributions as routing signals, TMoLE dynamically activates experts implemented with low‐rank adaptation to precisely capture trait representations and improve multi‐trait scoring. Experimental results on the constructed PTRE dataset show that TMoLE achieves an average Quadratic Weighted Kappa (QWK) of 0.6138. It achieves the highest average QWK among the compared methods and the best results on Textual Expression, Problem Awareness and Continuous Growth, supporting its effectiveness for content‐oriented multi‐trait scoring.
ABSTRACT Aiming at the problems of high missed detection rate and unbalanced feature extraction caused by the coexistence of multi‐scale objects in complex traffic scenarios, this paper proposes an improved YOLO11 detection algorithm (EPLS‐YOLO) for autonomous driving. First, the KITTI dataset is reconstructed and expanded, which is uniformly categorised into six classes (Car, Van, Truck, Pedestrian, Cyclist, and Tram). Data augmentation strategies such as Mosaic and random transformations are further adopted to enhance sample diversity. Second, in the backbone network, the Efficient Multi‐Scale Attention (EMA) is embedded into the C3K2 module to construct the C3K2_EMA module, which strengthens the fine‐grained feature representation of small objects and occluded objects. A P2 high‐resolution detection branch is added to cover the scale of extremely small objects, and a Lightweight Shared Convolution Detection Head (LSCD) is designed to achieve efficient fusion of multi‐scale features from shallow and deep layers. Finally, the SlideLoss function is introduced to dynamically assign sample learning weights based on Intersection over Union (IoU), alleviating the problem of unbalanced training of multi‐scale samples. Experimental results show that the proposed EPLS‐YOLO achieves a precision ( P ) of 94.5%, recall ( R ) of 92.0%, and mean Average Precision at IoU = 0.5 (mAP50) of 95.7% on the KITTI dataset, which are 0.7, 3.0 and 1.6 percentage‐points higher than those of the original YOLO11, respectively. Notably, the detection performance for small objects is significantly improved. Moreover, the overall detection performance of EPLS‐YOLO outperforms that of mainstream object detection models such as RT‐DETR, YOLOv8, YOLOv10 and YOLOv12, which can meet the demand for accurate perception of full‐scale objects in autonomous driving.
ABSTRACT With the rapid growth of e‐commerce, recommender systems have become essential tools for alleviating information overload by providing users with personalized item suggestions that match their preferences. Traditional collaborative filtering approaches have shown strong performance, while they still suffer from data sparsity since they rely solely on limited rating information. To address this problem, recent studies have incorporated user reviews as supplementary textual data, which contain rich semantic and emotional information reflecting user preferences and item characteristics. However, review texts often include noisy and irrelevant content that can degrade recommendation performance. To overcome this limitation, this study proposes a Summarization‐based Key Information‐Aware Recommendation (SKAR) model that integrates extractive summarization and multi‐embedding techniques to improve rating prediction accuracy. The proposed model first employs a TextRank‐based summarization module to extract key information from user and item reviews and reduce noise. It then applies a multi‐embedding module that fuses textual representations from BERT and RoBERTa to capture richer semantic features and reduce the representation bias arising from a single pretrained language model. Finally, a multi‐layer perceptron module predicts user ratings by modelling non‐linear user–item interactions. Experiments on three real‐world Amazon datasets show that the proposed SKAR model outperforms various baseline models. These results confirm the effectiveness of incorporating review summarization and multi‐embedding strategies in review‐based recommendation.
ABSTRACT Multimodal fake news detection (MFND) has attracted growing attention as misinformation increasingly appears in heterogeneous forms that combine text, images, audio, video, and social context, while recent generative models further increase the realism and scalability of deceptive content. Meanwhile, the rise of large language models (LLMs) and multimodal large language models (MLLMs) has introduced new opportunities as well as new evaluation challenges for MFND. In this survey, we present a structured review of the field through a taxonomy that groups existing methods into two broad paradigms: small‐model‐based methods and LLM‐involved methods. For each paradigm, we analyse representative approaches from the perspectives of modelling strategy, factual grounding, robustness, and deployment feasibility. We also review widely used datasets and evaluation protocols, and discuss their limitations with respect to modality composition, task heterogeneity, and result comparability. Finally, we outline important open problems and future directions, including the detection of LLM‐generated misinformation, cross‐lingual generalisation, interpretable and evidence‐grounded reasoning, and trustworthy evaluation in realistic deployment settings.
ABSTRACT Artificial Intelligence (AI) is increasingly utilised in medicine; however, its “black‐box” nature continues to hinder clinical trust, adoption, and validation. Explainable AI (XAI) has emerged as a critical field to address these transparency challenges by making AI‐driven decisions more interpretable and actionable. This narrative review examines the progress of XAI in medicine over the past decade. We first introduce fundamental XAI concepts and describe our review methodology, followed by a comprehensive analysis of key application domains, including medical imaging, electronic health records (EHRs), and multi‐omics data. Methodologically, we categorise XAI techniques into model‐agnostic approaches (e.g., SHAP, LIME, Anchors) and model‐specific approaches (e.g., Grad‐CAM, LRP, TreeSHAP). Beyond summarising their principles, advantages, and limitations, we further provide a systematic analysis of clinical reliability and failure modes associated with each class of methods, highlighting how explanation techniques may produce misleading, unstable, or non‐causal interpretations in real‐world clinical settings. The review then discusses the demonstrated benefits of XAI, including result validation, bias detection, and improved patient–clinician communication, while critically examining persistent challenges such as limited clinical deployment, inconsistent evaluation standards, and the lack of prospective validation. Finally, we outline future research directions, emphasising the need to adapt XAI to large‐scale foundation models and conversational AI systems, as well as to extend its applicability in biomedical and multi‐omics interpretation. We argue that while XAI is essential for improving transparency, trust, and clinical adoption, its reliable and scalable integration into clinical workflows requires both methodological innovation and rigorous, clinically grounded validation frameworks.
ABSTRACT The integration of large language models, particularly ChatGPT, into healthcare has attracted increasing attention due to their potential to support medical education, clinical decision‐making, administrative workflows, patient communication and clinical documentation. However, their adoption in healthcare also raises important concerns related to accuracy, reliability, privacy, professional supervision and validation in real clinical settings. This paper presents a systematic literature review (SLR) on the applications of ChatGPT in the health sector. The review follows established SLR guidelines and analyzes 25 primary studies selected from ACM Digital Library, IEEE Xplore and PubMed. The selected studies were examined according to four research questions focused on application areas, reported benefits, limitations and challenges and validation methods. The results show that ChatGPT has been mainly explored in medical education, diagnostic support, administrative task automation, oral health and medical research. Reported benefits include faster access to information, support for documentation, personalization, cost reduction and assistance in decision‐making. Nevertheless, the review also identifies relevant limitations, including hallucinations, inconsistent accuracy, context loss, ethical and legal risks, privacy concerns and dependence on prompt engineering. The analysis of validation methods shows that most studies rely on quantitative metrics, expert‐based assessment, qualitative evaluation or hybrid approaches, although real‐world clinical validation remains limited. This review provides a structured overview of current evidence and identifies research gaps that should be addressed to support the safe, reliable and effective use of ChatGPT and related LLM‐based tools in healthcare.
ABSTRACT Artificial intelligence (AI) is increasingly adopted in animal‐related domains such as health monitoring, behaviour analysis and welfare assessment. However, concerns about the transparency and interoperability of AI outputs are rising. This scoping review investigates how AI and explainable AI (XAI) are applied in animal‐related systems and examines the role of semantic technologies in enhancing their interoperability. In this review, we followed PRISMA‐ScR guidelines and conducted five structured searches across ScienceDirect, Springer, Scopus, IEEE Xplore and Web of Science. The searches targeted AI applications (S 1 ), XAI applications (S 2 ) and the use of semantic knowledge in AI (S 3 ) and XAI (S 4 ). Studies were screened, assessed using the QualSyst tool and selected based on Q1/Q2 SJR or CORE C–A* classification. A total of 21 review papers were selected for AI applications and 8 for XAI. No eligible papers were found regarding the use of semantics in AI or XAI. While explainability is gaining attention, it remains mostly limited to visual or statistical tools, lacking domain‐specific contextualization. We observed that AI is now pervasive in animal‐related research, yet XAI practices remain underdeveloped and lack semantic grounding. This gap calls for ontology‐based explainability to enhance trust, relevance and usability for both experts and non‐specialists.
ABSTRACT In recent years, there has been a rise in the use of ChatGPT for education, healthcare, smart cities and emerging technologies. However, no available studies or reviews have provided a consistent and reliable depiction of the situation regarding its usage and evaluation. The reporting of datasets, evaluation indicators, factors influencing performance and conditions of deployment has also varied from study to study. This fragmented state of affairs seriously inhibits attempts to assess ChatGPT's capabilities and limitations and thus improve the design of future versions. Earlier reviews were often conducted in a way pertaining to a single area or were mainly descriptive, with less emphasis on methodological evaluation and issues in deployment and ethics. To fill this void, we undertook a systematic review, according to PRISMA guidelines, limiting our searches to English‐language journal articles published during 2021–2025 by reputable publishers. Such studies focused directly on GPT models, provided assessment conditions and were cited extensively as preprints, while the exclusion criteria encompassed poorly linked studies, those not in English and studies employing ChatGPT as an adjunct. An analysis of these studies revealed that the vast majority of research has taken place in the area of education (32%) and health (28%). The review revealed significant variation in assessment accuracy across domains, frequent challenges with doubtful sensitivity, unpredictable and rapid changes and risks associated with specific domains that impact reliability and safety. This study's primary contribution is an effort to develop an integrated analytical framework that puts together these interdisciplinary results in a streamlined manner for interpreting the capabilities and limitations of ChatGPT. Because of the methodological heterogeneity of existing studies, the results can be viewed as qualitative trends instead of standard quantitative evidence. The results thereby accentuate the need for consistent criteria, domain‐informed evaluation practices and stronger methodological reporting to underpin a more reliable deployment of ChatGPT‐based systems.
ABSTRACT Accurate forecasting of agricultural prices is essential for informed production planning, market stabilisation and effective policy design. This review examines 773 studies published between 2006 and 2025 to synthesise recent advances. We begin by analysing the factor systems and structural characteristics of agri‐price data, and organise forecasting tasks by input–output design, temporal resolution and prediction objectives—ranging from point estimates to trend detection and probabilistic forecasting. Evaluation practices are reviewed across multiple dimensions, including error metrics, trend alignment, model selection and uncertainty estimation. We then trace the evolution of forecasting 15 methods from traditional statistical models to machine learning and deep neural 16 architectures (RNN, CNN, GNN, Transformer), as well as decomposition‐based and 17 ensemble strategies. These developments are contextualised within bibliometric trends, highlighting shifts in research focus and global collaboration. Empirical evidence shows that hybrid pipelines combining decomposition, feature learning and ensemble techniques tend to outperform standalone models, while simple linear models remain competitive for long‐horizon or low‐frequency forecasts. Common challenges include data leakage, inconsistent testing horizons and insufficient treatment of uncertainty. Looking ahead, future research should emphasise integrating diverse data sources—such as weather, trade and policy signals—and building models that can adapt to unexpected market changes. It is equally important to understand how price dynamics respond to policy actions, improve model transferability across regions and commodities and provide well‐calibrated forecasts with interpretable uncertainty estimates to enhance the practical value of agricultural price prediction.
ABSTRACT Federated learning (FL) has emerged as an impactful paradigm for privacy‐preserving machine learning, and allows model training without the need to share raw data. However, data heterogeneity across clients challenges practical FL deployment. Data space heterogeneity and statistical heterogeneity create significant training difficulties. System heterogeneity imposes additional external constraints. These combined factors impair convergence and reduce model performance. They also raise concerns regarding fairness, scalability and robustness. Focused on data heterogeneity, this review provides a structured analysis of FL. It encompasses three key areas: core categorizations of data heterogeneity, algorithmic advances (e.g., personalized FL, mixture‐of‐experts architectures, transfer learning‐based solutions) and system‐level techniques spanning communication optimization, resource adaptation and secure collaboration. We further synthesize benchmark efforts and real‐world applications in healthcare, finance, nuclear power and the Internet of Things (IoT)/edge computing to highlight the practical implications of heterogeneity‐aware FL. Finally, we identify key challenges and outline promising research directions towards scalable, fair and adaptive FL systems capable of operating in complex real‐world settings. This survey aims to serve as a reference point and conceptual roadmap for future research in heterogeneous FL.
ABSTRACT The growing global imperative for sustainable energy solutions is catalysing a transformative shift in materials science, where nanostructured materials, endowed with unique quantum and surface‐dependent properties, have emerged as critical drivers of next‐generation energy technologies. However, the extraordinary complexity of their design space makes traditional experimental discovery methods impractically slow, hindering progress towards urgent energy goals. The convergence of Artificial Intelligence (AI) and Machine Learning (ML) is revolutionising this landscape: AI/ML algorithms excel at rapidly screening vast virtual material libraries, predicting properties with high accuracy (often exceeding 90% for specific properties such as bandgap or stability), optimising nanostructures for targeted functionalities, and uncovering novel compositions via inverse design approaches. This review critically evaluates that AI‐driven methodologies, leveraging advanced tools such as graph neural networks and generative models, can drastically accelerate the design‐build‐test‐learn cycle, which is essential for addressing pressing global energy challenges. This review critically evaluates how AI‐driven methodologies align with expert‐system frameworks in materials science, emphasising decision‐support systems and rule‐based learning for accelerated discovery. Current applications highlight significant strides in key domains, including enhanced stability of perovskite photovoltaics, development of high‐performance solid‐state battery electrolytes, efficient catalysts for CO 2 reduction, and improved thermoelectric materials. This review also addresses persistent challenges, such as data scarcity, model interpretability, and the need for experimental validation. While challenges persist, such as data scarcity and the need for greater model interpretability, advances in materials informatics infrastructure and algorithmic sophistication underscore immense potential. This perspective analyzes AI's pivotal role in advancing the full potential of nanomaterials for a sustainable energy future, calling for strengthened interdisciplinary collaboration and integrated AI‐driven development to accelerate innovation. It also discusses limitations and conflicting findings in the current literature, providing a balanced assessment of the field.
ABSTRACT With the explosive growth in the volume of image usage, selective image encryption (SIE) has emerged as an efficient method to enhance encryption efficiency. The challenge of how to identify images containing sensitive content has long been a difficult issue. Deep learning technology, with its powerful semantic extraction capabilities, has naturally become an auxiliary tool for recognizing images containing specific content. This review primarily focuses on recent advancements in SIE integrated with semantic understanding. First, it reviews the current state of development in image ROI encryption. Subsequently, it proposes two semantic‐aware SIE schemes based on image‐to‐image and text‐to‐image search paradigms. The review also introduces evaluation metrics for assessing both the encryption algorithms and deep learning models involved in such SIE systems. Finally, it analyzes potential security issues in SIE, such as privacy protection of deep learning models and leakage of ROI edge regions, as well as possible optimization directions, including model lightweighting and encryption parallelization to enhance efficiency. In conclusion, this review indicates that selective encryption is not limited to ROI‐based approaches but also includes semantic retrieval followed by targeted encryption. Moreover, with the integration of deep learning models, considerations regarding security and efficiency have become more complex, representing key areas for further exploration in future research.
ABSTRACT The personalisation–privacy paradox captures the tension between using personal data for personalised services and respecting individuals' privacy. This study adopts a holistic research framework to clarify the paradox's core challenges, review existing approaches, and examine a user‐centric solution. First, this study presents a systematic review of research on the personalisation–privacy paradox, drawing on studies published in information systems, marketing and management journals. The review shows that current approaches often frame the paradox as a ‘dilemma,’ prioritising privacy at the expense of personalisation, rather than recognising it as a persistent tension between the two. We then review broader categories of privacy‐enhancing technologies (PETs), highlighting that user‐side, server‐side, and channel‐side solutions provide partial responses but rarely balance both demands. Against this backdrop, personal data stores (PDS) are examined as a hybrid approach that aims to give individuals greater control over their data while enabling more effective personalisation through selective sharing. Survey results indicate that privacy concerns are reduced when users better understand PDS, act rationally in weighing benefits and risks, and trust existing data protection regulations. These findings suggest that PDS have the potential to offer a synthetic response to the paradox by supporting both personalisation benefits and privacy protection.
ABSTRACT Artificial intelligence and machine learning (ML) now shape decisions in healthcare, finance and security, but they can reproduce historical prejudice and inequality. Bias in training data and in model implementation can amplify harm, especially for racial and gender minorities. Despite sustained research on fairness, mitigation in real‐world systems remains uneven, in part because stakeholders lack a shared and precise grasp of core notions, including bias, prejudice, discrimination and fairness. As a result, technical interventions are sometimes adopted without consistent conceptual grounding and reporting. This article addresses that problem by providing a knowledge base that aligns key concepts with empirical evidence and lifecycle stages. We conduct a scoping review to map sources of bias across the ML lifecycle and to identify forms of prejudice and discrimination associated with the use of sensitive attributes. We synthesize qualitative and quantitative evidence and introduce a conceptual model for organizing these findings. Our contributions are threefold: a refined lifecycle taxonomy of bias sources that introduces two additional types and spans all development stages; the explicit treatment of cognitive bias as a cross‐cutting meta‐bias; and an analysis of prejudice and discrimination that compiles a legally grounded catalogue of sensitive attributes and discusses their concepts and issues. Together, these results provide an integrated view of where and how bias emerges, and they support future research, evaluation and governance work on fairness in ML.
ABSTRACT Explainable AI (XAI) is crucial for fostering human trust in deep neural network (DNN) predictions, particularly in tasks like image classification. Multiple surveys exist on XAI methodologies, however, the practical usability and reproducibility of these methods remain largely unexplored. This paper addresses this gap by conducting a systematic survey of recent XAI papers published in leading computer vision and AI conferences and journals. We categorize these works, identify prevalent datasets and evaluation metrics, and analyse the associated code repositories. Our analysis reveals that almost 95% of the surveyed codebases are research prototypes rather than published releases, and a concerning majority of two‐thirds of them exhibit inconsistencies with their corresponding publications. These findings highlight the challenges in benchmarking new XAI methods against existing ones and explain the slow adoption of state‐of‐the‐art research in real‐world applications. This paper aims to underscore the importance of releasing well‐documented, readily usable code alongside XAI research to foster a more robust and reproducible ecosystem, ultimately facilitating the development and deployment of trustworthy AI systems. The results of this study are presented on an interactive website Interactive‐XAI.
ABSTRACT The effectiveness of Transformers in time series forecasting largely depends on their ability to extract and utilise key features. The model not only needs to capture temporal patterns but also focus its attention on the most informative inputs, thereby constructing sharp and discriminative attention distributions to highlight critical information and improve forecasting accuracy. However, despite their strengths in modelling global dependencies, conventional Transformers remain limited in capturing predictive key patterns and forming highly distinctive attention weights. To address this limitation, we propose FIRFormer, a framework that enhances input representation and reconstructs the attention mechanism to better capture important temporal dynamics. Specifically, we introduce the Dynamic Frequency Selector module to adaptively filter frequency components most relevant to the forecasting objective, reducing redundancy and improving input efficiency. Subsequently, we introduce the Biased Integration Window module, which integrates a temporal decay mechanism into sliding aggregation to strengthen the correlations within local contexts, achieves precise fusion of temporal dynamic patterns and further emphasises the modelling value of recent critical information. Finally, we propose the Attention Recalibration module, which employs a dual‐branch attention refinement strategy. It enhances attention selectivity from both the local feature perspective and the global subspace perspective, promoting sharper and more focused attention distributions. FIRFormer enhances the perception of critical dynamic patterns and achieves performance improvements across multiple real‐world forecasting tasks.
ABSTRACT Explainable AI (XAI) is essential for building trust in Deep Neural Networks (DNNs). SHAP (SHapley Additive exPlanations) is a well‐known XAI technique for attributing feature importance, but it struggles with exponential computational complexity as the number of features increases. Various approximation methods have been suggested, but they compromise SHAP's theoretical principles. We introduce AA‐SHAP, a novel approach that derives superpixel affinity from the explained model's internals to identify and group superpixels. AA‐SHAP constructs a relevance‐consistency affinity between superpixel interdependence, enabling much faster SHAP calculations on a reduced set of meta‐superpixels while outperforming previous methods in explanation faithfulness. Exact Shapley values are computed on the reduced meta‐superpixel game, preserving all axiomatic guarantees within the aggregated feature space. Evaluated across multiple datasets, both convolutional and transformer classification architectures show that AA‐SHAP produces more faithful attributions than competing methods while improving computational speed and maintaining SHAP's theoretical axioms. The source code is available at https://github.com/vhasic/AA‐SHAP .
