International Journal of Informatics and Data Science

Comparative Study of TF-IDF-SVM and IndoBERT for Imbalanced Indonesian Hoax Detection

2026-09-08T14:37:13+00:00

Indonesian hoax detection remains challenging because classification performance is influenced by text representation, class imbalance, and experimental design. This study comparatively evaluates a conventional Term Frequency–Inverse Document Frequency with Support Vector Machine (TF-IDF–SVM) approach and the transformer-based IndoBERT model under the same experimental protocol on a naturally imbalanced Indonesian hoax dataset. News titles and narratives were combined as textual input, duplicate records were removed, and the resulting dataset was stratified into training, validation, and held-out test sets. Model configurations were selected exclusively using validation Macro F1-score to prevent test-set information from influencing model selection. The final SVM employed unigram–bigram TF-IDF features with balanced LinearSVC, whereas the selected IndoBERT configuration used unweighted cross-entropy loss with a maximum sequence length of 128 tokens. On the test set, TF-IDF–SVM achieved an accuracy of 0.8299, Macro Precision of 0.7123, Macro Recall of 0.7065, and Macro F1-score of 0.7093, while IndoBERT achieved 0.8047, 0.6665, 0.6573, and 0.6616, respectively. The stronger numerical performance of TF-IDF–SVM suggests that discriminative lexical patterns remained highly informative in this dataset, while the contextual advantages of IndoBERT may not have been fully realized under the substantial class imbalance and the investigated training configuration. However, the exact McNemar test produced a p-value of 0.1371, indicating that the difference in paired classification correctness was not statistically significant at the 0.05 level. These findings demonstrate that TF-IDF–SVM remains a competitive and computationally simpler baseline for Indonesian hoax detection under imbalanced data conditions, while the benefit of more complex contextual models remains dependent on dataset characteristics and experimental settings.

Transfer Learning Performance and Efficiency of VGG16 and MobileNetV2 on Three-Class Animals-10 Classification

2026-09-04T12:06:33+00:00

Animal image classification remains challenging because variations in pose, illumination, object scale, viewing angle, and background can affect model predictions. In addition, high-capacity convolutional neural networks may impose substantial computational and storage requirements. This study compares the classification performance and computational efficiency of VGG16 and MobileNetV2 using transfer learning. Dog, Cat, and Sheep images were selected from the Animals-10 dataset to represent visually related companion animals and a visually distinct livestock category while allowing balanced sampling. Following image validation, duplicate checking, and class balancing, the final dataset contained 5,004 images, with 1,668 images per class. The dataset was divided using an identical stratified 80:20 split for both models, with 20% of the training portion used for validation. Both models used ImageNet weights, 224 × 224-pixel inputs, identical classification heads, data augmentation, and equivalent training configurations. MobileNetV2 achieved 97.80% accuracy, a 97.81% macro F1-score, and a 99.88% macro ROC-AUC, whereas VGG16 achieved 96.80%, 96.81%, and 99.67%, respectively. MobileNetV2 reduced training time by 66.03%, inference time by 59.56%, and weight size by 77.83%. However, the exact McNemar test indicated no statistically significant difference in classification performance (p = 0.064). Therefore, MobileNetV2 provided substantially greater computational efficiency while maintaining classification performance comparable to VGG16, making it a more practical option for resource-constrained animal image classification systems.

Comparative Study of Background Removal for Mango Fruit Disease Classification Using ResNet50-SVM

2026-09-04T11:55:36+00:00

Background removal is commonly applied in image classification to suppress irrelevant visual information and emphasize the target object. However, its contribution to disease classification may depend on whether background elimination also modifies contextual, boundary, and texture information used by deep feature extractors. This study evaluates the effect of background removal on five-class mango fruit disease classification using a hybrid ResNet50–Support Vector Machine (SVM) framework. A paired experimental design was constructed from 834 matched MangoFruitDDS image pairs representing Alternaria, Anthracnose, Black Mould Rot, Healthy, and Stem-End Rot. The paired data were stratified into 667 training and 167 test samples for each image condition. ImageNet-pretrained ResNet50 was used as a frozen feature extractor, producing 2,048-dimensional feature vectors that were standardized and classified using an optimized SVM. Background-removed images achieved a higher cross-validation macro F1-score (0.7559) than original images (0.7340). However, evaluation on the identical test set showed higher performance for original images, with 76.05% accuracy, 0.7687 macro F1-score, and 0.9541 macro ROC-AUC, compared with 74.25%, 0.7445, and 0.9425, respectively, after background removal. The effect was class-dependent: Healthy improved after background removal, whereas several disease classes declined, suggesting that removing the background may alter contextual, boundary, or texture-related information contributing to ResNet50 feature representations. Exact McNemar testing showed no statistically significant difference between the paired predictions (p = 0.783846). These findings indicate that developers should not assume hard background removal universally improves plant disease classification; class-specific validation and feature-preserving or attention-based alternatives should be considered.

Comparative Predictive Performance of Support Vector Regression and Random Forest in Forecasting CPO Spot Prices

2026-08-15T12:30:42+00:00

Accurate short-term forecasting of Crude Palm Oil (CPO) spot prices is critical for managing financial risk and stabilizing supply chains in volatile agricultural commodity markets. However, high non-linearity, temporal volatility, and sensitivity to market shocks present significant modeling challenges for traditional econometric tools. To address these issues, this study aims to develop an optimized machine learning framework that evaluates the comparative predictive performance of kernel-based margin regression against decision-tree ensembles for daily CPO spot price forecasting. Utilizing historical Malaysian CPO spot price data (2015–2024), the proposed solution combines multi-temporal feature engineering (autoregressive lags and simple moving averages) with a leak-free TimeSeriesSplit cross-validation and systematic grid-search hyperparameter tuning. Out-of-sample evaluation demonstrates that the tuned Support Vector Regression (SVR) model achieves superior predictive accuracy, registering an RMSE of 24.2191 USD, an MAE of 15.8937 USD, and a MAPE of 1.6940%. The SVR architecture significantly outperforms the optimized Random Forest (RF) Regressor (RMSE = 35.2395 USD, MAE = 23.8164 USD, MAPE = 2.5151%), yielding a 31.27% error reduction in RMSE. Feature importance analysis establishes that the 1-day lag (lag_1) contributes 66.75% of total split impurity reduction. Furthermore, SVR maintains homoscedastic stability during extreme late-2024 price surges exceeding 1113.00 USD, where Random Forest exhibits extrapolation truncation. This study contributes a validated, operationally sound decision-support blueprint for physical commodity risk management and clarifies algorithm selection under structural market shocks.

Beyond Vendor Hype: An Objective IDOCRIW-MARCOS Model for Selecting AI-Based Workforce Mental Health Platforms

2026-08-05T11:06:45+00:00

The growing adoption of artificial intelligence (AI) in workforce mental health monitoring has produced a wide range of competing platforms that differ in prediction accuracy, response time, scalability, implementation cost, and user satisfaction, so that HR and occupational-health decision-makers currently choose among them largely on the basis of vendor marketing claims rather than a structured, evidence-based comparison, making platform selection a genuine multi-criteria decision-making (MCDM) problem. This study proposes a decision support model that integrates the Integrated Determination of Objective Criteria Weights (IDOCRIW) method with the Measurement of Alternatives and Ranking according to Compromise Solution (MARCOS) method to give organizations an objective, reproducible alternative to subjective AHP/TOPSIS-style vendor evaluations for AI-based workforce mental health platforms. IDOCRIW combines Entropy and Criterion Impact Loss (CILOS) to derive objective criteria weights directly from vendor-reported technical specifications, removing the need for expert pairwise comparisons, while MARCOS ranks alternatives based on their utility degree relative to an ideal and an anti-ideal solution. The proposed model was demonstrated using an illustrative case study of five AI platform alternatives evaluated against five criteria: prediction accuracy, response time, scalability, implementation cost, and user satisfaction. The IDOCRIW results indicate that scalability (weight 0.260) is the most influential criterion, followed by response time (0.220). The MARCOS ranking identifies SmartWell as the top-ranked alternative with a final utility function value of 0.981, ahead of MentalCare AI (0.973) and WorkSense AI (0.969); a follow-up TOPSIS comparison computed on the same weighted data confirms SmartWell's top position but reverses the third- and fourth-ranked alternatives, showing that the ranking method itself, and not only the criteria weights, materially affects the recommendation. A sensitivity analysis, including combined-criterion perturbation scenarios, confirms that the ranking of the top alternative remains stable under moderate weight changes. For HR practitioners, the model offers a transparent and reproducible basis for AI-vendor procurement decisions that reduces reliance on unverified accuracy claims in vendor marketing materials.

Patient-Disjoint Evaluation for Trustworthy and Explainable Breast Cancer Histopathology Classification

2026-08-05T11:16:24+00:00

Breast cancer diagnosis from histopathological images increasingly relies on deep learning, with recent studies reporting classification accuracies above 95% on the benchmark BreaKHis dataset. However, much of this literature uses image-level splits that let the same patient's images appear in both training and test sets a source of inflation that is not merely statistical but potentially clinically misleading if taken as evidence of real-world reliability and rarely pairs high accuracy with a rigorous account of model interpretability. This study addresses that gap with an explainable deep learning pipeline for binary (benign/malignant) BreaKHis classification under a patient-disjoint split. An EfficientNetB0 backbone, pretrained on ImageNet, was fine-tuned in two phases with class-weighted loss and Reinhard-based stain normalization. Evaluated under this leakage-safe protocol, the baseline model achieved a test-set balanced accuracy of 0.765 markedly lower than the above-95% figures reported elsewhere, yet a more trustworthy generalization estimate together with an F1-score of 0.777 for the malignant class, a ROC-AUC of 0.849, and a PR-AUC of 0.940, with the best checkpoint selected before overfitting deepened during fine-tuning. A regularization ablation against a more heavily regularized variant showed a narrower generalization gap but lower test performance, confirming the unregularized configuration as the more suitable final model. Grad-CAM was applied to the model's final convolutional layer to visualize the regions driving individual predictions, supporting qualitative inspection of correct and misclassified cases. These findings argue for combining leakage-aware evaluation with explainability as a joint, rather than separate, requirement for trustworthy histopathological classification models.

Comparative Evaluation of YOLOv8 Variants for Static BISINDO Alphabet Detection: Accuracy and Computational Cost Analysis

2026-08-15T12:16:07+00:00

Indonesian Sign Language (BISINDO) plays an important role in communication for deaf communities in Indonesia. Automatic BISINDO alphabet recognition requires accurate detection while maintaining reasonable computational cost. This study evaluates the effect of model capacity on static BISINDO alphabet detection by comparing YOLOv8n, YOLOv8s, YOLOv8m, and YOLOv8l under consistent experimental conditions. A dataset of 11,468 images representing 26 alphabet classes was divided into 9,168 training, 1,155 validation, and 1,145 held-out test images. All models were trained for 50 epochs using 416 × 416 pixel images on an NVIDIA RTX 2060 with 6 GB memory. Performance was evaluated using precision, recall, mAP@0.5, mAP@0.5:0.95, and training time. The validation results achieved mAP@0.5 values of 99.42%, 99.46%, 99.38%, and 99.40% for YOLOv8n, YOLOv8s, YOLOv8m, and YOLOv8l, respectively, while training time increased from 0.93 to 3.59 hours. Held-out test evaluation remained consistently high, with the main errors involving the visually similar M and N gestures. Because the subsets originate from the same public dataset source, the results provide within-dataset evidence and do not establish cross-dataset or real-world generalization. Under the evaluated conditions, YOLOv8n provides the most efficient balance between detection performance and computational cost.