Journal of System and Computer Engineering

Attention-Driven Contrastive Learning for the Identification of Rare Partial Discharge Signal in GIS

Muhaimin Hading — 2026-07-29

Gas-insulated switchgear (GIS) is a critical component in high-voltage power transmission systems, where partial discharge (PD) activity can indicate early-stage insulation defects. However, phase-resolved partial discharge (PRPD)-based fault diagnosis remains challenging due to noisy signals, perturbed measurement conditions, and severe class imbalance, particularly for rare floating-electrode defects. This study proposes an attention-driven contrastive learning framework for rare PD signal identification in GIS. PRPD data are represented as two-dimensional density matrices derived from phase angle, discharge magnitude, and occurrence count. The proposed framework applies PRPD-specific data augmentation, followed by ResUNet-based denoising, CBAM-based feature refinement, and supervised contrastive learning to improve feature separability among PD classes. The framework was evaluated using a public 550 kV GIS PRPD dataset containing corona-type, surface-type, floating-electrode-type, and noise classes. The results show that augmentation substantially improved robustness. When trained with raw data, the proposed model achieved 91.90% accuracy and 53.81% F1-score under the original test scenario, but decreased to 48.76% accuracy and 46.26% F1-score under IEC-perturbed testing. After augmentation, the model achieved 98.24% accuracy and 96.58% F1-score under the original scenario, and maintained 97.44% accuracy and 97.76% F1-score under IEC perturbation. These findings indicate that the proposed framework supports robust PRPD representation learning for GIS PD diagnosis under perturbed and imbalanced conditions.

SIPETA Performance Evaluation Integration : EUCS and Equivalence Partitioning Technique

Litafira Syahadiyanti — 2026-07-29

The development of information systems in higher education institutions requires comprehensive performance evaluations to ensure that the systems are capable of supporting learning and administrative processes effectively, efficiently, and in accordance with user needs. This study aims to evaluate the performance of SIPETA in supporting academic processes and to identify factors that affect the quality of the system. The main problems found were suboptimal information quality, data accuracy, and several system functions that affected user trust. The study used an integrative approach with the End User Computing Satisfaction (EUCS) model to measure user satisfaction and the Equivalence Partitioning technique in Black Box testing to assess system functionality. Data was collected through questionnaires administered to 100 users and direct testing of the system's features. The results showed that the system had an effectiveness rate of 86 percent and was rated as good in terms of appearance, ease of use, and timeliness of service. However, the content and accuracy variables were still in the poor category due to several functional failures such as schedule validation, notifications, revision uploads, and guidance history. Overall, SIPETA is suitable for use but requires improvements in data quality and system logic to increase reliability and user satisfaction.

Comparison of SMOTE, Class Weighting, and Classical Machine Learning Models on the ID-SMSA Indonesian Stock Market Dataset

I Komang Adyanata — 2026-07-30

Sentiment classification of social-media text related to the Indonesian stock market is a growing research area. The ID-SMSA dataset is the publicly available labelled corpus for this domain, yet class-imbalance handling strategies on this dataset have not been systematically compared across multiple classifiers. This paper evaluates Multinomial Naive Bayes, linear Support Vector Machine (SVM), and Random Forest under three imbalance-handling conditions: no handling, class weighting, and SMOTE. All experiments use the full 3,287-tweet dataset with an 80:20 stratified split and report macro F1 as the primary metric. SMOTE consistently improves macro F1 across all classifiers. The largest gain is on Naive Bayes (+0.137, from 0.589 to 0.726). The best configuration is SVM with SMOTE, achieving macro F1 of 0.752 and accuracy of 0.784. Class weighting benefits Random Forest (+0.011) but slightly reduces SVM, confirming that linear SVM on TF-IDF is robust to moderate imbalance at IR = 2.41. Per-issuer evaluation reveals macro F1 variation from 0.647 on TPIA to 0.881 on BBNI, shaped by vocabulary consistency, class dominance, and domain specificity. These results provide a transparent and reproducible classical baseline that situates transformer-based and deep-learning approaches on ID-SMSA within a well-defined reference frame.

Explainable Non-Organic Waste Classification: A Comparative Study of CNN with SHAP Interpretability Against Vision Transformer Approaches

Setio Basuki — 2026-07-30

This study develops a classification model based on Machine Learning (ML) and Computer Vision (CV) to automatically distinguish recyclable and non-recyclable waste. Waste management still faces the challenge of low manual sorting efficiency, thus reducing recycling potential. The dataset consists of 399 waste images divided with a ratio of 80:10:10 for training, validation, and testing. Three combinations of visual features were tested, namely Mean Color (RGB) + LBP + HOG, Histogram + HOG + Edge, and Mean Color (RGB) + Histogram + GLCM, each of which was evaluated using four conventional ML algorithms, namely SVM-RBF, SVM-Linear, Random Forest, and Gradient Boosting. Meanwhile, deep learning models namely CNN, ViT, and LoRA ViT were trained directly on raw images without manual feature extraction. Experimental results show that CNN achieved the highest testing accuracy of 82.50%, outperforming all conventional ML models that achieved a maximum accuracy of 75.00%, as well as ViT (72.50%) and LoRA ViT (70.00%). The application of SHAP-based Explainable AI (XAI) provides transparency to the model's decision-making process. These findings demonstrate that CNN with certain regularization settings are effective for distinguishing recyclable and non-recyclable waste, in supporting sustainable smart waste management systems.

Development Of Deep Learning in Diagnosing Pathology

Januardi Nasir — 2026-07-30

Skin cancer is one of the most common types of cancer worldwide, with a particularly high incidence in Indonesia. According to Globocan 2020 data, there were approximately 18,000 cases of skin cancer with nearly 3,000 deaths, with melanoma having the highest mortality rate. A study at Dr. Cipto Mangunkusumo General Hospital (2014–2017) showed that malignant melanoma accounted for 5.7% of all skin cancer cases, with the majority of patients presenting at an advanced stage. Pathological diagnosis remains the gold standard for confirming melanocyte lesions, but it is subjective with variability reaching 45.5%.The development of Whole Slide Imaging (WSI) and Deep Learning (DL) has enabled the implementation of more accurate and consistent computer-aided pathological diagnosis systems. Several previous studies, such as those by Hekler et al., Brinker et al., and Li et al., have demonstrated that Convolutional Neural Network (CNN)-based models can match or exceed the performance of human pathologists. However, two major challenges remain: the model's limitations in distinguishing atypical melanocytic lesions and decreased performance due to staining variations across medical centers. This study aims to develop a DL-based intelligent pathological diagnosis model using WSI images that can accurately distinguish benign, atypical, and malignant melanocytic lesions and is robust to staining variations, to improve the effectiveness of skin cancer diagnosis in Indonesia.

Performance Comparison of YOLOv8 and YOLOv11 for Manga Text Area Detection Based on the Intersection over Union Metric

M Ridwan Dwi Septian — 2026-07-30

Manga has complex visual characteristics, such as variations in speech bubble shapes, diverse text orientations, and dense background illustrations, which complicate the automatic text area detection process. Detection errors in the form of false positives and false negatives can cause text areas to be localized inaccurately and affect the processing at subsequent stages. This study compares the performance of YOLOv8m and YOLOv11m in detecting text areas in Japanese manga using the Intersection over Union (IoU) metric. The dataset consists of 551 manga images annotated into three classes, namely clean_text, messy_text, and text_bubble. Both models were trained under the same parameter configuration for 60 epochs to ensure an objective comparison. The evaluation was performed on 50 test images covering 564 text objects. The test results show that YOLOv11m obtained an average IoU of 0.7598, which is higher than YOLOv8m (0.7196). In addition, YOLOv11m exhibited a faster inference time of 1158.72ms compared with 1276.23ms for YOLOv8m. Based on these results, YOLOv11m demonstrated superior performance over YOLOv8m in terms of both localization accuracy and computational efficiency for the Japanese manga text area detection task.

Aspect-Based Sentiment Analysis of YouTube Comments on CoreTax Using Support Vector Machine and Random Forest

Nur Vadila — 2026-07-30

The implementation of the CoreTax Administration System (CTAS) by the Directorate General of Taxes has received diverse responses from the public, which are reflected in YouTube comments. This study applies Aspect-Based Sentiment Analysis (ABSA) to identify user opinions regarding CoreTax and compares the performance of Support Vector Machine (SVM) and Random Forest for sentiment classification. Data were collected through web scraping from three Youtube videos, yielding 1.527 valid comments after preprocessing. A rule-based method was used to classify comments into five aspects, namely system, performance, user-friendliness, tax services, and policy. The results indicate that the system aspect was the most frequently discussed (56,12%), while negative sentiment dominated the dataset (59,2%). The highest proportion of negative sentiment was found in the user-friendliness aspect (81,29%), followed by performance (76,44%). In model evaluation, Random Forest achieved better results than SVM, obtaining 0.80 accuracy, 0.84 precision, 0.75 recall, and 0.79 F1-Score. Overall, ABSA provides deeper insights into user perceptions and issues related to CoreTax implementation.

Explainable Machine Learning for Long-Term Monthly Hydroclimatic Forecasting and Extreme-Event Detection

Arif Fadillah — 2026-07-30

Long-term hydroclimatic prediction in arid urban environments remains methodologically demanding because monthly records are often intermittent, highly seasonal, zero-inflated, and dominated by rare but consequential extreme events. Using a 121-year monthly hydroclimatic record for Makkah, Saudi Arabia, spanning January 1901 to December 2021, this study develops an explainable hybrid machine-learning framework for monthly forecasting, seasonal diagnostics, and extreme-event detection. The dataset contains 1,452 monthly observations with a mean value of 6.19, median of 3.00, standard deviation of 8.05, and maximum of 52.00, indicating a strongly skewed distribution. Exploratory analysis reveals pronounced seasonality: November, December, and January exhibit the highest hydroclimatic values, whereas June is consistently dry across the full record. A temporal feature set was constructed using lag variables, rolling statistics, annual seasonal memory, cyclical month encodings, and trend indicators. Several predictive models were evaluated, including Random Forest, Extra Trees, Histogram Gradient Boosting, XGBoost, and a hybrid SARIMA–Random Forest residual-correction model. Extra Trees achieved the best forecasting performance on the holdout period, with MAE = 2.997, RMSE = 5.603, sMAPE = 57.669%, and R² = 0.518. Extreme-event detection was performed using a 90th-percentile threshold of 17.68, identifying 146 extreme months over the full record. The best classification trade-off was obtained by Histogram Gradient Boosting, while Random Forest produced the highest ROC-AUC. SHAP-based interpretation demonstrates that seasonal phase variables and annual memory features dominate model behaviour, especially month_cos, month_sin, same_month_last_year, and lag_12. The findings show that interpretable ensemble learning can provide a more transparent and operationally relevant framework than accuracy-only forecasting for arid-region hydroclimatic risk assessment.

Public Approval Rating Updates for President Prabowo Using a Bayesian Dirichlet-Multinomial Hybrid Model

anwar anwar anwar — 2026-07-30

Conventional estimation of presidential approval ratings relies on face-to-face surveys that excel in population representativeness (probability sampling) but suffer from significant time lag between data collection and publication. Conversely, social media monitoring (SMM) provides real-time data but is often biased toward active digital platform demographics. This study proposes a hybrid model based on Bayesian Dirichlet-Multinomial (BDM) to integrate survey data as prior belief with SMM data as likelihood to generate more dynamic and accurate posterior estimates. Prior data were drawn from the Indikator Politik Indonesia national survey of October 2025 (n=1,220, MoE ±2.9%), indicating a public approval rate of 77.7%. Update data were collected from the X (Twitter) platform during November 2025 through crawling and scraping techniques, yielding 2.5 million raw data points processed through Named Entity Recognition (NER), text normalization, and Transformer-based classification (IndoBERT v2.0). After deduplication, 310,000 unique accounts were retained with 120 million total impressions. Sentiment distribution showed 61% positive, 15% negative, and 24% neutral. Bayesian updating produced a corrected public approval estimate of 68.4% (a 9.3 percentage point decline from the survey prior), detecting pockets of dissatisfaction on food price issues underrepresented in conventional surveys. Model validation using Leave-One-Out Cross-Validation (LOO-CV) produced an Expected Log-Pointwise Predictive Density (ELPD) of -127.4, superior to the single-survey baseline model (-148.9). This model offers a new framework as a public opinion early warning system that is responsive to current issue dynamics without sacrificing the statistical validity of traditional survey methods.

Computer Vision-Based Plastic Waste Detection and Weight Estimation

Andri Dwi Utomo — 2026-07-30

Plastic waste management is an increasingly critical environmental challenge due to the growing volume of waste and the limitations of conventional manual weighing methods, which are inefficient and prone to human error. This study proposes a computer vision-based system for automatic detection, classification, and weight estimation of plastic waste using the YOLOv8n model for object detection and Random Forest Regression for weight estimation. The YOLOv8n model is used to detect and classify seven types of plastic waste based on the Resin Identification Code (RIC), namely PET, HDPE, PVC, LDPE, PP, PS, and OTHER. Subsequently, weight estimation is performed using a Random Forest Regression model based on bounding box features, including width, height, area, aspect ratio, and perimeter. The proposed system is evaluated using an unseen test set to ensure unbiased performance measurement. Experimental results show that the YOLOv8n model achieves a mean Average Precision (mAP@0.5) of 91.93% and mAP@0.5:0.95 of 73.27%, while the Random Forest Regression model achieves an R² score of 95.5% with a Mean Absolute Error (MAE) of 4.28 grams. These results demonstrate that the integration of object detection and regression enables accurate and automatic estimation of plastic waste weight, thereby improving the efficiency and objectivity of waste management systems.

DNN vs CNN for YouTube Shorts Abusive Word Classification

Evan Febditya Pratama — 2026-07-30

Abusive words in YouTube Shorts are difficult to moderate because harmful expressions may appear in captions, comments, or embedded text within visual frames. This study compares the performance of a Deep Neural Network (DNN) and a Convolutional Neural Network (CNN) for abusive-word classification using text and image features from YouTube Shorts content. The dataset consisted of 1,013 image-text samples labeled into two binary classes: abusive and non-abusive. Textual features were derived from the text column representing OCR/caption information and transformed into 2,500-dimensional features, while image features were obtained from RGB images resized to 96 by 96 pixels. The dataset was divided into training, validation, and testing subsets consisting of 658, 202, and 153 samples, respectively. Both models were trained using dropout, L2 regularization, early stopping, and image augmentation. The results show that CNN outperformed DNN, achieving test accuracy of 0.8627 and macro F1-score of 0.8527, compared with DNN accuracy of 0.8301 and macro F1-score of 0.8195. These findings indicate that CNN provides more effective multimodal classification performance for detecting abusive words in YouTube Shorts content.