International Journal of Advances in Artificial Intelligence and Machine Learning

Learning from Heterogeneous Label Quality: Strategies for Robust Model Training

Wu Shukun, Hafiz Muhammad Kurniawan, Adolf Asih Suprianto — Mon, 28 Sep 2026 00:00:00 +0700

Background: Training data collected from crowdsourcing platforms, automated labeling systems, and multiple annotators often contain labels with unequal levels of reliability. Conventional machine-learning methods generally treat all labels as equally trustworthy, which can reduce model performance when label errors vary across samples and annotation sources.
Aims: This study proposes a quality-aware training method that explicitly models the heterogeneity of label reliability. It examines whether sample-level quality weighting can improve predictive performance and robustness under synthetic and real-world label-noise conditions.
Methods: The proposed method combines annotation or neighborhood agreement, loss-based confidence, and source-level reliability to estimate a continuous quality score for each training sample. The scores are incorporated into a weighted cross-entropy objective and periodically updated during training. Experiments were conducted on CIFAR-10, CIFAR-100, CIFAR-10N, and CIFAR-100N using ResNet-18. The method was compared with cross-entropy, label smoothing, generalized cross-entropy, Co-teaching, and DivideMix.
Result: Under high heterogeneous noise, the proposed method achieved accuracies of 83.21% on CIFAR-10 and 54.62% on CIFAR-100, compared with 73.48% and 45.92% for conventional cross-entropy. It also achieved the highest accuracy under the evaluated real-world noisy-label conditions. Ablation and sensitivity analyses showed that the combined quality indicators were complementary and that the method remained effective under moderate quality-estimation errors.
Conclusion: Explicitly modeling label reliability improves robustness when training data contain unequal annotation quality. The proposed method provides a practical and computationally efficient strategy for learning from crowdsourced, weakly supervised, and multi-source datasets.

Out of Distribution Detection for Reliable Deployment of Machine Learning Models

William Yeoh, Fadly Fadly, Maria Ulfa — Mon, 28 Sep 2026 00:00:00 +0700

Background: Machine learning models can give confident predictions when they receive inputs that are different from the data used during training. These predictions may not always be reliable when the model encounters unfamiliar inputs. OOD detection can help identify such inputs and reduce the risk of incorrect predictions in real-world applications.
Aims: This study evaluates several OOD detection methods and examines their ability to reduce unreliable automatic predictions.
Methods: Several post hoc detection methods were evaluated using ResNet classifiers trained on CIFAR datasets. The methods included Maximum Softmax Probability, ODIN, Energy-Based Detection, Mahalanobis Distance, Deep k Nearest Neighbors, and ReAct. The evaluation covered near-OOD, far-OOD, and corrupted-in-distribution conditions. Detection performance was also examined together with prediction errors, review needs, processing time, and memory use.
Results: The methods showed different performance across the tested conditions. Deep k-Nearest Neighbors performed well for near OOD inputs, while ReAct showed strong performance for far OOD detection. The deployment simulation showed that ReAct-based gating reduced OOD leakage and improved the accuracy of predictions that were processed automatically. It also reduced incorrect predictions made without additional review. Energy-based detection provided competitive results with relatively low computational requirements.
Conclusion: OOD detection can improve the reliability of machine learning systems. When combined with abstention and fallback mechanisms, unfamiliar inputs can be identified and sent for further review. This approach reduces unreliable predictions while allowing familiar inputs to continue through the automated process.

Adaptive Curriculum Learning for Accelerating Convergence in Data-Driven Model Training

Eko Risdianto, Muhammet Esad Kuloğlu, Maria Ulfa — Mon, 28 Sep 2026 00:00:00 +0700

Background: Efficient training of machine learning models remains a critical challenge in data-driven artificial intelligence systems, particularly when training samples vary in difficulty, noise level, and representational complexity.
Aims: This study proposes and evaluates an adaptive curriculum learning framework that accelerates convergence in machine learning model training by dynamically adjusting sample ordering based on the model's evolving competence.
Methods: The proposed approach introduces a feedback-driven curriculum mechanism that integrates difficulty estimation and model competence evaluation. Training samples are initially ranked by estimated difficulty and are progressively updated during training using validation-based performance signals. Experiments are conducted on three benchmark datasets, including UCI Adult Income, Breast Cancer Wisconsin, and CIFAR-10, using multilayer perceptron and convolutional neural network architectures. The framework is compared against random sampling, static curriculum learning, and hard-first sampling strategies. Performance is evaluated based on convergence speed, predictive accuracy, robustness to noise, class imbalance handling, and computational overhead.
Result: The proposed adaptive curriculum learning framework consistently accelerates convergence across all datasets compared to baseline methods. The model achieves faster stabilization of training dynamics while maintaining or slightly improving predictive performance on accuracy, F1-score, and AUC. Furthermore, the framework demonstrates improved robustness under noisy-label conditions and class-imbalance scenarios, with reduced performance degradation compared to conventional sampling strategies. The computational overhead introduced by the adaptive mechanism remains minimal compared with the gains in training efficiency.
Conclusion: The study concludes that aligning sample ordering with evolving model competence significantly improves training efficiency in machine learning systems.

Graph-Based Machine Learning for Modeling Complex Relationships in High-Dimensional Data

Sri Huning Anwariningsih, Lusiana Efrizoni, Misinem Misinem — Sun, 27 Sep 2026 00:00:00 +0700

Background: High-dimensional datasets frequently encode complex relational dependencies that are inadequately captured by conventional vector-based machine learning models operating under independent and identically distributed assumptions.
Aims: This study aims to develop an end-to-end graph-based machine learning pipeline that systematically converts high-dimensional structured data into relational graph representations.
Methods: The proposed methodology constructs relational graphs from raw feature spaces using similarity-based and interaction-driven edge formation strategies. The framework integrates scalable graph neural network training with neighborhood sampling techniques and structural attribution analysis. Comparative benchmarking is conducted against strong non-relational baselines, including multilayer perceptrons, Random Forests, and gradient boosting models. Evaluation employs standardized train validation–test splits, cross-validation protocols, and statistical significance testing across multiple heterogeneous datasets.
Result: Experimental findings show that graph-based architectures consistently outperform non-relational baselines, achieving improvements of 6–12% in F1-score and ROC-AUC across diverse datasets. Graph Attention Networks demonstrate enhanced robustness under noisy conditions and provide interpretable neighbor importance distributions. Scalability experiments confirm that sampling-based training preserves computational feasibility for large-scale graphs without substantial performance degradation.
Conclusion: The results demonstrate that incorporating relational inductive bias through graph modeling significantly enhances predictive accuracy, robustness, and structural interpretability in high-dimensional environments.

Energy-Efficient Artificial Intelligence: Evaluating Sustainable Model Design in Data-Intensive Systems

Deshinta Arrova Dewi, Firdhaus Hari Saputra Al Haris — Sun, 27 Sep 2026 00:00:00 +0700

Background: The rapid growth of artificial intelligence (AI) in data-intensive systems has significantly increased computational demand, raising concerns regarding energy consumption and environmental sustainability.
Aims: This study aims to develop a unified benchmarking framework that integrates predictive performance, computational complexity, and real-time energy consumption into a sustainability-aware evaluation protocol.
Methods: The proposed framework evaluates multiple model families by jointly measuring accuracy metrics, computational complexity, and real-time energy consumption during training and inference. Experiments were conducted under controlled hardware and dataset configurations to ensure comparability. Trade-off analysis was performed using sustainability-adjusted efficiency metrics and Pareto-optimal frontier evaluation to assess energy-performance balance.
Result: Experimental results show that deep neural networks achieved the highest predictive accuracy, with improvements of up to 2.1% over ensemble-based baselines. However, these gains required substantially higher energy consumption during training and inference. Gradient Boosting models demonstrated the most favorable sustainability-adjusted efficiency, reducing total energy usage by approximately 35–45% compared to deep architectures while maintaining competitive accuracy. Lightweight neural networks further achieved an optimal balance, occupying the Pareto-efficient frontier in energy-performance trade-off analysis.
Conclusion: The findings confirm that marginal accuracy improvements often incur disproportionate energy costs, underscoring the importance of sustainability-integrated benchmarking frameworks.

Self-Supervised Learning Strategies for Reducing Label Dependency in Large-Scale Data Science Applications

Kit Ling Chan, Mohd Zaki Zakaria, Eka Puji Agustini — Sun, 27 Sep 2026 00:00:00 +0700

Background: The rapid expansion of large-scale data ecosystems has intensified the challenge of acquiring high-quality labeled datasets for supervised learning models. Although raw data availability continues to grow exponentially, manual annotation remains costly, time-consuming, and susceptible to inconsistency, creating a structural bottleneck in scalable machine learning deployment. Existing self-supervised learning (SSL) approaches have demonstrated promising results in domain-specific applications, particularly in computer vision. However, many lack a unified, modality-agnostic framework and systematic evaluation under realistic constraints such as limited-label availability and noisy annotations.
Aims: This study aims to develop a unified self-supervised learning framework that reduces label dependency across heterogeneous data modalities. The scope of the research includes image and tabular datasets, focusing on evaluating model robustness under low-label conditions and label noise scenarios.
Methods: The proposed framework integrates contrastive representation learning with reconstruction-based objectives in a two-stage training strategy. First, large-scale unlabeled pre-training is conducted to learn robust feature representations. This is followed by supervised fine-tuning under controlled label fractions (1%–20%) and synthetic label corruption levels of up to 30%. Performance is evaluated against fully supervised baselines using accuracy and data efficiency metrics.
Result: Experimental findings demonstrate that the proposed SSL framework achieves up to 15% higher accuracy than fully supervised baselines under 1% labeled data conditions. Moreover, it maintains approximately 85–90% of clean-label performance even with 30% label corruption. The data efficiency ratio improves by up to 2.3× in low-label regimes, indicating substantial gains in learning effectiveness under constrained supervision.
Conclusion: The results confirm that self-supervised pre-training significantly enhances representation quality, robustness, and scalability while reducing annotation requirements. This research establishes SSL as a foundational paradigm for developing data-efficient and resilient AI systems in large-scale analytics environments.

Constraint-Aware Machine Learning for Ensuring Feasible Predictions in Operational Data Science

Lau Meng Cheng, Adolf Asih Suprianto — Wed, 23 Sep 2026 00:00:00 +0700

Background: Large-scale machine learning models require substantial computational resources during training, particularly in GPU-intensive and distributed computing environments. Although modern training pipelines achieve high predictive performance, they often rely on static resource allocation strategies that do not adapt to evolving learning dynamics. This leads to inefficient GPU utilization, increased training time, and unnecessary computational cost, limiting the scalability and sustainability of large-scale AI systems.
Aims: This study aims to improve training efficiency by proposing a Dynamic Resource Allocation (DRA) framework that integrates real-time learning signals into computational resource management. The framework dynamically adjusts GPU allocation based on convergence behavior, enabling efficient alignment between computational demand and model training stages.
Methods: The proposed framework employs a descriptive, analytical, and comparative experimental design, using secondary, confidential machine learning training logs comprising 1,500 training runs. The system integrates a training-monitoring module, a convergence-analysis mechanism, and an adaptive resource controller. Performance evaluation is conducted by comparing static allocation and dynamic allocation strategies using metrics such as training time, GPU utilization, model accuracy, and computational efficiency.
Results: Experimental results demonstrate that the proposed framework significantly improves training efficiency. The dynamic allocation strategy reduces training time by approximately 30–32%, increases GPU utilization by up to 17%, and improves overall computational efficiency without degrading model accuracy. Furthermore, the convergence analysis shows that the proposed method achieves faster, more stable convergence than static allocation strategies.
Conclusion: The findings confirm that integrating training-aware resource allocation into machine learning pipelines significantly enhances both efficiency and sustainability. By dynamically aligning computational resources with learning behavior, the proposed framework reduces wasteful computation while maintaining predictive performance. This approach provides a scalable solution for efficient large-scale model training in cloud and high-performance computing environments.

Fault Warning Using Clustering Analysis Method on Wind Turbine Blade

Yuli Guo, Bing Zeng; Yirui Qiao — Wed, 09 Sep 2026 00:00:00 +0700

Background: In an attempt to solve the problem of internal damage and crack of wind turbine blades that are difficult to be evaluated, a method based on cluster analysis is proposed to study the fault warning of wind turbine blades using machine learning.
Aims: Developing an early warning method (fault warning) for explicit damage (including surface cracks and internal damage) to wind turbine blades by applying cluster analysis (K-Means) and unsupervised learning to short-term historical operational data.
Method: Firstly, 2 MW wind turbine was used to collect short-term historical operation data of wind turbine and perform data pre-processing. Secondly, the wind speed-power, wind speed-hub speed and wind speed-tip speed ratio were used to classify the parameters affecting the explicit faults of wind turbine blades by using the cluster analysis method, and the distance analysis was performed by using the combined weight method segmentation, and the axial vibration data of the abnormal data points were analyzed by using the Fourier series transformation. And then, establishing the wind turbine blade explicit fault evaluation rules based on the performance reliability theory. Lastly, the feasibility of the method is verified by cases.
Result: The results demonstrated that the method can quickly warning the wind turbine blade explicit faults, especially for evaluation of the wind turbine blade surface explicit cracks and internal explicit fault.
Conclusion: This work has achieved the fault warning of large components based on the wind turbine short-term historical operation data.

AI-Driven Adaptive Online Digital Modules for Communication Courses Using Learning Analytics and Natural Language Processing

Fri, 21 Aug 2026 00:00:00 +0700

Background: The rapid growth of online learning in higher education requires innovative solutions that support independent, scalable, and standardized learning experiences. Artificial Intelligence (AI), particularly Learning Analytics (LA) and Natural Language Processing (NLP), offers opportunities to enhance digital learning through adaptive content delivery, personalized learning pathways, and automated formative feedback.
Aims: This study aimed to develop and evaluate an AI-driven adaptive online digital module for Communication courses that supports personalized learning and standardized instruction across higher education institutions using Gemini.
Methods: This study employed the ADDIE development model, comprising Analysis, Design, Development, Implementation, and Evaluation. Learning Analytics was used to monitor student engagement and learning progress, while NLP analyzed students' written responses to generate automated formative feedback. The module was validated by instructional design, subject matter, and media experts, followed by individual and small-group trials. Its effectiveness was evaluated using normalized gain (N-Gain) analysis.
Results: Expert validation, individual trials, and small-group evaluations indicated that the developed module achieved a "very good" level of feasibility. The effectiveness evaluation produced a high N-Gain score (0.7), indicating a substantial improvement in student learning outcomes. The integration of Learning Analytics and NLP supported adaptive learning, timely feedback, and increased student engagement.
Conclusion: The AI-driven adaptive digital module is feasible and effective for supporting online learning in Communication courses. Integrating Learning Analytics and Natural Language Processing enables personalized instruction and data-informed learning support, making the module a promising approach for improving learning quality in higher education.