Machine learning (ML) has become an important data-driven approach for classification and prediction, including applications in medical diagnosis. Ensemble methods can improve classifier performance by combining complementary learning mechanisms. However, systematic evidence on the sequential use of Bagging and XGBoost across different classifier architectures remains limited. This study develops a staged ensemble framework in which five classifiers—Support Vector Machine (SVM), Neural Network (NN), Logistic Regression (LR), Decision Tree (DT), and K-Nearest Neighbours (KNN)—are first optimized through Bagging and subsequently enhanced using XGBoost. The experiments were conducted on the Breast Cancer Wisconsin (Diagnostic) dataset under a consistent 70:30 train–test protocol. Performance was assessed using accuracy, confusion matrices, ROC curves, and Area Under the Curve (AUC), while repeated experiments were used to examine statistical significance. The results show that the staged Bagging–XGBoost approach improves both predictive accuracy and class discrimination across the evaluated classifier types. Neural Network achieved the largest improvement, with mean accuracy increasing from 93.1% to 97.0% across repeated experiments. The findings demonstrate that the sequential framework can improve non-tree-based as well as tree-based classifiers, providing empirical evidence for broader use of staged ensemble integration in breast cancer classification.
The rapid growth of generative artificial intelligence has increased the spread of deephoax images, creating significant challenges for digital security, public trust, and the reliability of online information. This study evaluates the performance of MobileNet and Xception architectures for classifying facial images into fact and hoax categories while analyzing different feature representation levels within the Xception architecture through an ablation-based fine-tuning strategy. The proposed framework consists of data preparation, model architecture design, training procedures, and performance evaluation using confusion matrix analysis and metrics such as accuracy, precision, recall, and F1-score. Two publicly available datasets, DeepDetect2025 and FF-GenAI were utilized to evaluate the robustness of the proposed models. Experimental results show that Xception consistently outperformed MobileNet across all evaluation metrics, with the Middle Level Layer configuration achieving the best performance at 99.71% accuracy and F1-score. The findings indicate that intermediate feature representations are the most effective for capturing structural inconsistencies, texture irregularities, and synthetic blending artifacts commonly found in AI-generated facial images. In contrast, low-level representations were less discriminative, while highly abstract semantic representations slightly reduced sensitivity to localized manipulation artifacts. Overall, this study demonstrates the effectiveness of Xception-based feature refinement for deephoax image detection and contributes to AI-based approaches for digital content verification, cybersecurity, and visual misinformation mitigation.
Public sentiment analysis regarding government policies, particularly in niche sectors like Halal tourism, often faces significant computational challenges due to data scarcity and severe class imbalance. Traditional augmentation methods, such as Synthetic Minority Over-sampling Technique (SMOTE) applied to raw text or TF-IDF vectors, often degrade semantic integrity, while standard fine-tuning of pre-trained models like IndoBERT tends to bias predictions toward majority classes. To address these limitations, this study proposes IndoBERT-LSS (Latent Space Sampling), a novel decoupled two-stage deep learning framework. The first stage employs a Representation Learning approach integrating an Attention-Pooling mechanism with Supervised Contrastive Loss (SupCon) to enforce compact intra-class clustering in the embedding space. The second stage introduces a Latent Space Sampling strategy, where SMOTE is applied to the extracted high-dimensional embeddings rather than the raw text, followed by a final classification using a Multi-Layer Perceptron. Validated on a dataset of 1,051 textual responses regarding Halal tourism in Pariaman, Indonesia, the proposed model achieved an accuracy of 93.84% and a macro F1-score of 0.94. Notably, the model demonstrated exceptional robustness in identifying minority classes, achieving a 0.99 F1-score for neutral sentiments. These results conclude that decoupling representation learning from feature balancing in the latent space significantly enhances model performance on imbalanced short-text datasets compared to standard fine-tuning baselines.
Polycystic Ovary Syndrome (PCOS) is commonly assessed with ovarian ultrasonography, but speckle can conceal follicular margins and reduce the robustness of automated interpretation. Although automated PCOS studies increasingly employ machine learning, the contribution of conventional despeckling to subsequent segmentation and classification has not been examined consistently. This study compares five classical filters - Mean, Median, Lee, Frost, and Kuan - within an interpretable machine-learning pipeline for ovarian ultrasound analysis. From a public collection of 12,680 images, a balanced sample of 300 scans (150 PCOS and 150 non-PCOS) was selected. Two radiologists produced follicle annotations, and disagreements were resolved with a third expert to obtain consensus masks. Each filtered image was segmented by adaptive thresholding with morphological refinement, after which geometric and intensity descriptors were extracted. Support Vector Machine (SVM), Random Forest (RF), k-Nearest Neighbors (k-NN), and Logistic Regression (LR) were trained using a stratified 70/30 train-test split with cross-validated hyperparameter tuning. The Kuan-LR configuration yielded the strongest result, reaching 94.44% accuracy and an AUC of 0.98, together with the best edge-preservation score and segmentation agreement. The results indicate that preprocessing materially affects the reliability of an interpretable PCOS detection pipeline and provide quantitative guidance for selecting a speckle-reduction strategy before segmentation and classification.
The purpose of the study was to develop an effective approach to automated generation and validation of educational assessment materials in the context of distance learning using Large Language Models (LLM). While current research often overlooks the dynamic transformation of evaluation practices using sequential AI applications, this study addresses this gap by introducing a novel self-correcting AI pipeline that iteratively refines and validates questions based on structured feedback. As part of the study, an experimental methodology combined qualitative and quantitative methods, including computer modelling. Researchers conducted a comparative analysis of the results from three different LLMs: ChatGPT-4o, Google Gemini PRO 1.5, and Claude 3.5 Sonnet. Finally, the generated questions were validated based on pedagogical expediency, age compliance, and Bloom’s taxonomy. To test the effectiveness of the approach, three stages of research were conducted. In the first stage, it was found that only 23% of questions generated without updated prompts met the stated criteria. In the second stage, after introducing clarifications to the prompts, this indicator increased to 63%. The highest results were achieved in the third stage, where an iterative hint refinement model using structured feedback was implemented: Claude 3.5 Sonnet achieved 92% of valid questions with the minimum number of clarifications (16), ChatGPT-4o 80% (40 clarifications), and Gemini 72%, but with the highest number of corrections (108). The process included question generation, answer verification, external validation, and iterative correction. The findings showed that effective AI-based development of multi-level computer science questions requires high-quality language models, clear instructions, and automated verification of cognitive level. The proposed algorithm enables educational institutions, platform developers, and teachers to generate assessment materials adapted to students’ age and knowledge levels in distance or blended learning
