Machine learning models are increasingly deployed in high-stakes decision systems where predictive performance alone may provide an insufficient basis for responsible model selection. Although explainable artificial intelligence (XAI) offers techniques for interpreting model behavior, existing evaluation approaches often assess explanation properties independently and provide limited support for integrated model selection. This study proposes the Explainability–Governance Score (EGS) Framework, a multidimensional decision-support methodology integrating four complementary dimensions: Fidelity, local predictive Stability, Counterfactual Actionability, and Governance Alignment. The framework combines established XAI techniques, practical recourse, and governance-oriented criteria into a composite evaluation of candidate models. It is empirically evaluated using five machine learning models across three high-stakes domains represented by the German Credit, Pima Indians Diabetes, and COMPAS datasets. Under the primary BPS specification, predictive evaluation ranked XGBoost, Decision Tree, and Random Forest highest in Credit Risk, Healthcare, and Criminal Justice, respectively. Under the multidimensional EGS evaluation, Logistic Regression achieved the highest EGS in Credit Risk (0.915) and Healthcare (0.915), while Decision Tree achieved the highest EGS in Criminal Justice (0.886). The identity of the highest-EGS model remained unchanged across all six prespecified weighting configurations in each domain. These findings demonstrate that multidimensional EGS evaluation can materially influence candidate-model rankings relative to predictive evaluation and provide a structured, reproducible basis for evaluating candidate models beyond predictive performance alone.
Whole-slide images and computed tomography scans are commonly decomposed into patches or slices, whereas diagnostic labels are assigned only at the study level, so that which patches or slices support the diagnosis remains unknown. Multiple-instance learning addresses this weakly supervised setting through aggregation functions, which determine how instance-level evidence produces the bag-level prediction. Several probabilistic MIL methods rely on max-based aggregation, which retains only the strongest instance evidence. On the other hand, the standard noisy-OR combines all instance scores, so that the bag probability depends on the product between the bag cardinality and the average instance activation. As a consequence, a slide covered by thousands of benign patches can receive the same probability as one containing a small tumor region. We introduce Correlated Chained Gaussian Processes with Weighted Noisy-OR Likelihood (CCGP-WNOR), a framework in which each instance is assigned a weight that modulates its contribution to the bag probability, with the weights summing to one so that the aggregation is normalized. Consequently, the bag probability lies between the minimum and maximum instance scores, and a value close to one requires at least one instance carrying strong positive evidence. Instance scores and aggregation weights are obtained by transforming two latent functions coupled through a semiparametric latent factor model with shared Gaussian processes, permitting cross-output dependence in the prior. Posterior uncertainty is propagated through a factorized approximation that preserves marginal means and variances but omits cross-output and cross-instance covariances from likelihood integration and prediction. Scalable inference is performed using inducing variables and stochastic variational inference. Experiments on a semi-synthetic benchmark and three medical imaging datasets show that CCGP-WNOR achieves a competitive performance among the compared state-of-the-art MIL methods.
Despite recent advances in multilingual artificial intelligence, Indigenous languages of the Americas remain underrepresented in natural language processing (NLP) because of limited corpora, annotations, benchmarks, and reusable resources. This systematic review examines AI-based NLP research on low-resource Indigenous languages published within the 2022–2026 review window. Following PRISMA 2020 guidance and an expanded source-verification and supplementary-search process, the final analytical corpus comprised 283 unique records. Machine translation was the most represented application domain (58.3%), followed by language documentation (51.2%), Indigenous-language education (30.4%), and speech technologies (28.6%). Transformer-related architectures (54.1%) and BERT-based models (34.6%) showed the greatest literature visibility, although verified implementation/evaluation was substantially lower than terminology-level occurrence. Quechua (39.2%) and Guarani (33.6%) were the most represented languages. Multilingual training and fine-tuning were the most frequently identified adaptation strategies, while Accuracy, BLEU, and chrF predominated among evaluation metrics. Overall, the evidence reveals a developing but structurally uneven and methodologically heterogeneous field. Persistent challenges include data scarcity, linguistic complexity, orthographic and dialectal variation, limited benchmark availability, uneven community participation, and underreported Indigenous data-governance practices.
Physics-informed neural networks (PINNs) approximate solutions of differential equations by embedding governing equations and boundary conditions into the training objective. Controlled PINN comparisons are difficult because geometry representation, collocation sampling, loss construction, neural architecture, and optimization settings can each affect the result. This paper presents MultiPINN, a modular PyTorch framework for component-level PINN research. MultiPINN decomposes the PINN workflow into composable geometries, boundary entities, physical conditions, point generators, neural models, loss terms, and training services. It supports analytical and mesh-based geometries, constructive operations, boundary-shell abstractions with normal-vector handling, condition-specific non-adaptive and residual-adaptive sampling, loss regularization, mixed-precision training, gradient accumulation, distributed multi-GPU execution, and callback-based monitoring. We evaluate the framework on representative benchmarks covering geometry handling, sampling strategies, neural architectures, three-dimensional nonlinear simulations, and computational performance. The results show that MultiPINN enables controlled studies of individual PINN components while preserving a unified implementation workflow. We also compare MultiPINN with existing PINN frameworks and position it as a compact research environment for experiments with complex geometries and configurable training pipelines.
Electricity theft is a worldwide challenge that causes significant losses for utility companies, with annual revenue losses estimated at USD 89.3 billion globally. The widespread rollout of Advanced Metering Infrastructure (AMI) and smart meters has made it easier to access large amounts of high-frequency consumption data from consumers. This access allows companies to use data-driven strategies for electricity theft detection. This review looks at how well machine learning (ML) and deep learning (DL) models detect electricity theft (ETD) using three benchmark datasets. The datasets presented are the State Grid Corporation of China (SGCC) dataset, the Open Energy Data Initiative Theft Detection Dataset 2022 (OEDI/TDD2022), and the Irish Commission for Energy Regulation Smart Metering Project (CER SMP) dataset. A thorough evaluation of 27 models is conducted, covering classical machine learning classifiers, tree ensemble methods, anomaly detection techniques, recurrent neural networks, convolutional networks, Transformer-based encoders, and federated learning systems. Key preprocessing methods, including filling in missing values, normalization, feature engineering, and tackling class imbalance with SMOTE and LoRAS, are noted for each model. The findings show that deep learning methods consistently outperform traditional machine learning techniques across all three benchmark datasets. The CNN-LSTM model, improved with LoRAS, achieves the top reported F1-score of 97.7% on the SGCC dataset. Meanwhile, the Federated ConvGRU model reaches the highest accuracy of 97.8% on OEDI/TDD2022. Transformer architectures show the best overall performance in average categories, with a mean accuracy of 95.8% and a mean F1-score of 95.5%. Addressing class imbalance is seen as the most influential factor, showing an 18.7 percentage-point difference in F1 scores between unaugmented and augmented LSTM models on SGCC. This review concludes that combining hybrid augmentation-driven deep learning architectures with privacy-preserving federated learning frameworks offers the most scalable and practical way to design Electricity theft detection (ETD) systems. The research community should prioritize standardized cross-dataset benchmarking and mandatory reporting of computational complexity as immediate priorities.
