Meta learning is a promising paradigm in the era of large models, and task distributional robustness has become an indispensable consideration in real-world scenarios. Recent advances have examined the effectiveness of tail task risk minimization in fast adaptation robustness improvement. This work contributes to more theoretical investigations and practical enhancements in the field. Specifically, we reduce the distributionally robust strategy to a max-min optimization problem, constitute the Stackelberg equilibrium as the solution concept, and estimate the convergence rate. Under certain scenarios, we incorporate the diversity regularizer into the acquisition criteria design during active subset selection and further improve meta learners' comprehensive generalization under tail risk minimization. In the presence of tail risk, we further derive the generalization bound, establish connections with estimated quantiles, systematically analyze the diversity regularizer's impacts, and practically improve the studied strategy. Accordingly, extensive evaluations on tasks such as few-shot sinusoid regression, system identification, image classification, and meta reinforcement learning, along with experiments on multimodal large models, demonstrate the significance, robustness and scalability of our proposal.

BackgroundAccurate patient mortality prediction is crucial in the emergency department (ED) to improve emergency healthcare services. Current prediction models are limited in both accuracy and practicality, particularly in identifying high-risk patients early. Machine Learning is deeply affected by data quality because mortality samples are significantly fewer than survival samples. This study aimed to achieve balanced and better accuracy of patient mortality prediction and evaluate the effectiveness of data balancing methods.MethodsThis study analyzed 2,437,341 non-traumatic adult ED visit records collected between 2008 and 2016 from five medical centers in Taiwan, including four mortality timeframes: death within 24, 72, and 168 h, and final death, and evaluated three data balancing methods: Random Under Sampling (RUS), Synthesized Minority Oversampling Technique (SMOTE), and Random Over Sampling (ROS). We adopted Random Forest (RF), AdaBoost (ADA), XG Boost (XGB). Logistic Regression (LR) is the meta learner for these models. Besides, we performed feature importance analysis based on RF, ADA, AdaBoost with BootStrap (ADA-BS), and Information Gain (IG).ResultsOur model with XGB achieved the best AUROC, 91.41%, which is better than 90.2% in the previous study by Wu et al. using the same dataset in 168-hour mortality timeframe. Our True Positive Rate (TPR) and True Negative Rate (TNR) are 79.88% and 86.73%, which are more balanced than 25% and 100% in the previous study. ROS achieves the better results than RUS and SMOTE and becomes our primary data balancing method. While adopting XGB in 24-hour mortality timeframes, ROS achieved the best AUROC, 93.72%, RUS achieved 93.61% and SMOTE achieved 91.73%. Compared with the previous study by Lin et al., the feature importance analysis shows our balanced dataset has better feature importance impacts, especially for the “Age” and “Triage” features.ConclusionOur method achieves better AUROC than the previous study, especially in the long challenging death-hour mortality timeframe with XGB and ROS. Our method achieves balanced TPR and TNR, which are more practical than AUROC. Besides, feature importance analysis shows our balanced dataset has better feature importance impacts.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12911-026-03608-9.

An ensemble learning approach for predicting hospital stay in transplant patients

Ecological impacts of hydropower projects in the lower reaches of the Yarlung Zangbo River driven by machine learning: Focusing on nitrogen and phosphorus cycle prediction and assessment

Accurate retail sales forecasting is crucial for understanding customer demands, handling inventories, and optimizing business strategies. Conventional forecasting approaches struggle to take into consideration both linear and nonlinear transformations in the time series data, limiting the model's adaptability. Additionally, an effective business strategy is essential for improving overall company revenue, depending on insights from precise forecasting. In order to address these shortcomings, the Meta-Learning Enhanced Learnable Long Short-Term Memory network (Meta-LLSTM) is proposed for effective retail sales forecasting and generalization. The model applies meta learning for quickly adapting with enhanced forecasting under diverse operational conditions. On the contrary, the Multiple-Parameter Exponential Linear Unit (MPELU) introduces learnable parameters, enabling the model to handle both linear and nonlinear transformations to effectively forecast retail sales in business strategies. To improve retail sales forecasting, the following support modules are used in this study: Recency, Frequency, Monetary and Diversity (RFMD) to analyze customer sales information, K-means based customer segmentation, and Adaptive Inventory Correction (AIC) for information on inventories. Additionally, the metrics of Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Coefficient of Determination (R-squared), Mean Absolute Percentage Error (MAPE) and Symmetric Mean Absolute Percentage Error (SMAPE) are used to evaluate the Meta-LLSTM. The proposed model achieves an RMSE of 1.003 which is less than that of the state-of-the-art classifiers such as Auto Encoder (AE), Convolutional Neural Network (CNN), Gated Recurrent Unit (GRU) and Recurrent Neural Network (RNN). Specifically, the RMSE of Meta-LLSTM is 16.97% less than the state of art approach (RNN), rendering it more effective than the existing models.

Missing data is a challenge in clinical research, especially in real-world data (RWD), where complete case analysis can bias results and reduce power. Ensemble learning approaches like Super Learner (SL) show strong numerical performance for prediction problems, but their use for missing value imputation (MVI) in oncology datasets is unexplored. We sought to develop and evaluate a novel SL-based imputation function that can impute multiple variables and quantify observation-specific uncertainty. We analyzed two independent cohorts of acute myeloid leukemia patients (n = 1641), 546 patients from the University of Colorado and 1095 from an external real-world cohort. The SL-based MVI function includes data processing, predictor selection, binary and continuous variable pipelines, and automatic performance measurement. Ensembles for both binary and continuous variables integrate diverse base learners, such as generalized linear models, random forests, and neural networks, via a meta learner that optimizes predictive accuracy. The binary variable pipeline's SL ensemble was optimized using area under the curve (AUC), while the continuous variable pipeline's SL ensemble was optimized via non-negative least squares. Performance was compared to multiple imputation by chained equations (MICE) using balanced accuracy, F1-score, root mean square error (RMSE), and visualizations. Observation-specific uncertainty was quantified for all imputations of both binary and continuous variables, with both additionally having lower and upper resampling-based potential imputation values. The SL cross-validation loop, SL ensemble trained for imputation, and resampling all supported parallelization. Clinically significant features of the cohorts were selected a priori based on prior literature. In a numerical experiment with 9 clinically important binary features, the proposed MVI function imputed and achieved higher balanced accuracy than MICE for 7/9 variables (mean balanced accuracy 89.04% vs. 80.75%) with comparable performance for the other 2 variables. The continuous variable SL ensemble, across 4 variables, showed an average 24.45% lower RMSE than MICE. On average, the SL ensemble trained for prediction took 145.02s to process for binary targets. This study demonstrates that the SL-based imputation function has improved performance over MICE in high-dimensional RWD while providing novel, observation-level uncertainty quantification.

Stacked ensemble machine learning for phenological wheat yield prediction from UAV-RGB and multispectral satellite data

Meta learning recently has been heavily researched and helped advance the contemporary machine learning. However, achieving well-performing meta-learning model requires a large amount of training tasks with high-quality meta-data representing the underlying task generalization goal, which is sometimes difficult and expensive to obtain for real applications. Current meta-data-driven meta-learning approaches, however, are fairly hard to train satisfactory meta-models with imperfect training tasks. To address this issue, we suggest a meta-knowledge informed meta-learning (MKIML) framework to improve meta-learning by additionally integrating compensated meta-knowledge into meta-learning process. We preliminarily integrate task-agnostic meta-knowledge into meta-objective via using an appropriate meta-regularization (MR) objective to regularize capacity complexity of the meta-model function class to facilitate better generalization on unseen tasks. As a practical implementation, we introduce data augmentation consistency to encode invariance as meta-knowledge for instantiating MR objective, denoted by DAC-MR. The proposed DAC-MR is hopeful to learn well-performing meta-models from training tasks with noisy, sparse or unavailable meta-data. We theoretically demonstrate that DAC-MR can be treated as a proxy meta-objective used to evaluate meta-model without high-quality meta-data. Besides, meta-data-driven meta-loss objective combined with DAC-MR is capable of achieving better meta-level generalization. 12 meta-learning tasks with different network architectures and benchmarks substantiate the capability of our DAC-MR on aiding meta-model learning. Fine performance of DAC-MR are obtained across all settings, and are well-aligned with our theoretical insights. This implies that our DAC-MR is problem-agnostic, and hopeful to be readily applied to extensive meta-learning problems and tasks. All codes for reproducing our experimental results are released at https://github.com/xjtushujun/DAC-MR.

Evaluation of a global repository of published dairy cattle methane emission prediction models on a northern irish dataset: Individual performance vs a stacked ensemble approach.

EETTE: Efficient evolutionary travel time estimation with deep meta learning

Abstract Bearing fault diagnosis is critical to industrial equipment safety, but in real-world scenarios it often suffers from extremely scarce fault samples and severe cross-domain distribution shifts, which seriously limit model generalization. To this end, this paper integrates data augmentation and cross-domain diagnosis and proposes a few-shot cross-domain bearing fault diagnosis method, DADM-PCML, based on a Domain-Adversarial Diffusion Model (DADM) and Poly-Contrastive Meta Learning (PCML). For data augmentation, we build a DADM and couple a Domain Discriminator Module (DDM) with the reverse-process training, where prior knowledge of the real data distribution is used to explicitly constrain generation and reduce the domain discrepancy between generated and original data. To further address insufficient generation diversity and unstable training, we design a staged freezing mechanism and a progressive alternating training strategy that dynamically adjusts the adversarial update frequency to balance the training process. For cross-domain diagnosis, to tackle pronounced cross-domain distribution discrepancies and chaotic feature distributions, we develop a meta-learning-based diagnostic method, PCML. By jointly imposing classification and contrastive loss constraints, PCML improves the model's feature discriminability while maintaining cross-domain diagnostic accuracy. Finally, the abundant data generated by DADM are used to provide PCML with sufficient training support. Experiments on the CWRU and rail transit bearing datasets show that the proposed method achieves 96.11% and 99.00% diagnostic accuracy, respectively, in extremely few-shot cross-domain scenarios.

There are a number of heutagogical strategies that assist students control their own learning, reflect on what they have learned and how they have learnt it, and boost their self-efficacy.Additionally, heutagogy encourages student cooperation. Self-study, textbook learning, study reference books, internet searches, group discussions for fresh ideas, cooperative learning, collaborative learning, mobile learning, meta learning, cognitive learning, library study, assistance from parents, tutors, and society, communication with teachers for clarification of doubts, etc. are examples of heutagogical practices.

Repeat prostate biopsy prediction remains limited by small patient cohorts that constrain artificial intelligence application despite theoretical advantages in capturing complex clinical patterns. This study develops and validates a meta-learning optimized TabNet framework using readily available clinical parameters to overcome sample size constraints and enhance repeat biopsy (RB) prediction accuracy through knowledge transfer from larger initial biopsy (IB) cohorts, with particular applicability to resource-limited settings where mpMRI remains unavailable. Meta-learning enables rapid model adaptation by leveraging knowledge from related tasks with minimal training examples. This retrospective study analyzed 2,087 initial prostate biopsies and 139 subsequent RBs without mpMRI data. A two-stage training paradigm implemented Model-Agnostic Meta-Learning for pre-training on IB data, followed by fine-tuning on the RB cohort. Performance evaluation included discrimination analysis, calibration assessment, and decision curve analysis compared to original TabNet and conventional machine learning approaches, with classification performance benchmarked against established clinical risk calculators. Among 139 RB patients, cancer was detected in 40 cases (28.8%), including 31 clinically significant cancers (75.5%). On the independent testing set of 42 patients, meta-learning TabNet achieved superior discriminative performance (AUROC 0.872) compared to XGBoost (0.808), original TabNet (0.800), and conventional approaches. The model demonstrated optimal calibration (Brier score 0.068, ECE 0.100) and high specificity (90.0%) with only three false positives, substantially outperforming ERSPC and PCPT calculators. Meta-learning optimization successfully addresses sample size limitations in repeat prostate biopsy prediction without requiring advanced imaging. This provides an evidence-based decision support tool enhancing diagnostic accuracy while minimizing unnecessary procedures.

Accurate and robust prediction of drug-target affinity (DTA) plays a critical role in drug discovery. While deep learning has advanced DTA prediction, existing methods struggle with limited training data and poor generalization. In this study, we propose AdaMBind, a novel DTA prediction model based on meta-learning framework with an adaptive task module designed for low-data scenarios. It employs a dynamic "easy-to-hard" task scheduling mechanism to enhance training efficiency and robustness. Experimental results on three benchmark datasets demonstrate that AdaMBind outperforms 8 baseline models in predicting affinity for unseen targets, particularly under few-shot conditions. Under stringent data constraints, the model successfully identifies high-affinity compounds for ESR and TP53, achieving outstanding virtual screening performance. Furthermore, when applied to inhibitor discovery against FLT3 for acute myeloid leukemia, AdaMBind successfully identified candidate compounds with potent inhibitory activity, as verified by preliminary experimental assays. In summary, AdaMBind provides a robust framework for few-shot DTA prediction.

Supply Chain Management (SCM) is an essential component of any successful enterprise’s supply chain. For a business’s continued operation and positive reputation, its well-organized product production and shipping processes are significant. With all the progress deep reinforcement learning has made in logistics, there are still models that cannot handle relational data in supply chain graphs or adapt quickly enough to new sustainability regulations. To bridge the gap, a Sustainability-aware Hierarchical Graph-driven Meta Reinforcement Learning Optimizer (SHG-MRLO) is proposed, integrating hierarchical policy architectures, Graph Neural Network (GNN), and meta learning for scalable adaptation and improved performance. The optimizer model supplies a supply chain as a dynamic graph, where a meta controller dynamically orchestrates decision strategies across multilevel graph-embedded states, ensuring robust adaptation to unexpected disruptions and demanding sustainability constraints. An experimental evaluation on benchmark datasets simulating a realistic supply chain scenario demonstrates that the proposed model achieves a 7.5% increase in on-time delivery, a 28.6% reduction in carbon emissions, and a path coverage rate of 93.3%, outperforming the baseline models. The presented SHG-MRLO technique can independently achieve SCM policies in the face of a complex, adaptive environment. This work contributes a scalable and interpretable framework for enhancing sustainable supply chain performance, merging the robustness of meta-learning with the expressiveness of graph-based modelling, and paves the way for the next generation of eco-efficient supply chain solutions.

ABSTRACT Landslides are devastating natural disasters, resulting in loss of life and property. We can reduce the impact of landslides by identifying the susceptible areas and regularly monitoring slope instability. This article presents a novel stacked ensemble machine learning (ML) model for landslide susceptibility mapping in Meghalaya, India. The proposed stacking model consists of the base layer and the meta layer. The base layer employed Logistic Regression, XGBoost, Random Forest, LightGBM, Support Vector Classifier, Extra Trees, k-Nearest Neighbours, Decision Tree, Gaussian Naive Bayes, and Quadratic Discriminant Analysis as the base learners. A neural network meta learner then processes the intermediate features (meta-features) to predict the probability of landslide occurrence. Recursive feature elimination with cross validation and the variance inflation factor were employed to determine the optimum features. Sixteen landslide causative factors, along with 2070 landslide and non-landslide points, were utilised to train and evaluate the proposed model. The proposed model's performance was assessed using accuracy, AUC, F1-score, precision, recall, and Kappa. It outperformed the individual base learners and advanced ML methods, including deep learning and hybrid approaches. Finally, the proposed model was employed for generating a landslide susceptibility map of Meghalaya.

With the rapid development of artificial intelligence, knowledge graphs have become increasingly important for many downstream reasoning tasks. Although large-scale knowledge graphs have been constructed in many domains, many relations still suffer from data sparsity, making few-shot knowledge graph completion a practical and challenging problem. Existing methods typically enhance entity embeddings by aggregating one-hop neighbors. However, these neighbors often contain many irrelevant entities to the target relation, which introduces a large amount of noise. And existing methods are not sensitive to the semantic differences and task characteristics in the reference set, which leads to a lack of deep semantic meanings in relation representations and weakens relation inference. To address these challenges, we propose meta learning based few shot knowledge graph completion with domain selected aggregation. Specifically, the method introduces a domain-selected neighborhood aggregation mechanism that dynamically filters irrelevant neighbors through a selection strategy and a gating module, effectively suppressing noise propagation under sparse data conditions. Moreover, a relation meta-learner is designed by integrating contextual attention with a multi-layer perceptron to capture deep semantic correlations among reference triples and generate more expressive task-aware relational representations. An embedding learner further utilizes a meta-optimization strategy to enable rapid adaptation to new tasks. Experiments on NELL-One and Wiki-One datasets demonstrate that the method significantly outperforms state-of-the-art baselines. Specifically, in 5-shot tasks, it achieves performance improvements of [Formula: see text] and [Formula: see text] on the Hits@10 metric for datasets, respectively.

Abstract Global climate change has brought a paradigm shift in groundwater management practices. For a sustainable groundwater supply, there is a need to predict the availability of groundwater due to the intermittent fluctuations in the discharge and recharge phenomena. This study proposes an optimised stacking generalisation methodology for groundwater level prediction. The proposed methodology had the Particle Swarm Optimization-Artificial Neural Network (PSO-ANN), Genetic Algorithm-Artificial Neural Network (GA-ANN) and Self-Adaptive Differential Evolutionary Extreme Learning Machine (SaDE-ELM) as the base learners. Intercomparison revealed the prediction strength of SaDE-ELM over the other base learners. Hence, the stacked model was formed when the SaDE-ELM was used as the meta learner. Statistical analyses show that the proposed stacking method performed better than the standalone hybrid PSO-ANN, GA-ANN and SaDE-ELM with average RMSE and R values of 0.2209 m and 0.78, respectively. The superiority of the stacking method was further revealed using the Taylor diagram to present the statistical comparison with the observations of all models used.

MetaGT: A lightweight graph transformer via meta learning for large-scale graphs

Accurately estimating state of health (SOH) for lithium-ion batteries based on machine learning methods usually entails large requirement of training data, bringing difficulties for practical applications. To address this challenge, this study proposes a novel SOH estimation method integrating meta-learning, temporal convolutional networks (TCNs), and transformers with limited sampling data. First, by dividing constant current charging curves into multiple segments, the capacity increment sequences for each segment are extracted as health features. A parallel hybrid network is developed, which combines the strengths of TCNs, transformers, and attention mechanisms to effectively capture both local and global patterns in health features. In addition, meta-learning is employed with small sampling retraining data to improve the model adaptability acrossvarying temperatures, different charging currents and battery chemistries. Experimental validations conducted on different temperatures and charging currents show that the proposed method achieves the maximum estimation error of 3%. Moreover, when applied to different types of batteries, the proposed method requires only a small amount of target battery data for retraining to achieve performance comparable to traditional methods, thereby reducing the need for aging data. These results underscore the robust generalizability, high accuracy, and strong potential for real-world applications of the proposed method.