Jurnal Sains, Teknologi & Komputer

Optimization of bilstm accuracy for gold price forecasting

2025-10-13T23:47:37+07:00

Gold price forecasting is challenging in the financial sector due to its volatile and nonlinear movements, which are influenced by various economic factors. This study aims to implement and optimize a Bidirectional Long Short-Term Memory (BiLSTM) model for forecasting weekly gold prices. The data used consist of weekly world gold closing prices from 2020 to 2024 obtained from Yahoo Finance. The data preprocessing stages include normalization using Min-Max Scaling, sliding window construction, and data splitting into training and testing sets. Model optimization was performed by evaluating combinations of hyperparameters, including the number of neurons, epochs, batch size, and dropout. Model performance was evaluated using Root Mean Squared Error (RMSE) and Mean Absolute Percentage Error (MAPE). The experimental results show that the optimal hyperparameter configuration achieved an RMSE of 83.99 and a MAPE of 2.41%. These results indicate that the BiLSTM model is capable of capturing the weekly gold price movement patterns in the dataset and can be used as an approach for deep learning-based price forecasting.

Comparative analysis of convolutional neural network architectures for guava disease classification

2025-09-25T13:37:52+07:00

This study aims to compare the performance of two Convolutional Neural Network (CNN) architectures, EfficientNetB0 and EfficientNetV2-S, in classifying guava (Psidium guajava L.) conditions based on digital images. The dataset consists of three classes: anthracnose, fruit fly infestation, and healthy fruits. Class imbalance was addressed using an undersampling technique, while both models were developed using a transfer learning approach with ImageNet pre-trained weights and trained for 30 epochs. Model performance was evaluated using accuracy, precision, recall, and F1-score. The results showed that EfficientNetB0 achieved a validation accuracy of 99.49% with a loss of 0.0124, whereas EfficientNetV2-S achieved a validation accuracy of 94.17% with a loss of 0.2153. The classification report further showed that EfficientNetB0 achieved F1-scores ranging from 0.99 to 1.00 across the three classes, while EfficientNetV2-S achieved F1-scores ranging from 0.92 to 0.96. Based on these results, EfficientNetB0 demonstrated higher classification performance than EfficientNetV2-S on the dataset and training configuration used in this study.

Artificial intelligence governance and digital leadership as predictors of organizational effectiveness in public secondary schools and primary healthcare centres: evidence from Anambra State, Nigeria

2026-08-11T08:46:51+07:00

This study examined artificial intelligence (AI) governance and digital leadership as predictors of organizational effectiveness in public secondary schools and primary healthcare centres in Anambra State, Nigeria. A correlational research design was adopted, targeting a census population of 609 administrators, of whom 590 returned completed questionnaires. Data were collected using three researcher-developed questionnaires on AI governance, digital leadership, and organizational effectiveness, with reliability coefficients ranging from 0.84 to 0.88. Simple and multiple linear regression analyses were used for data analysis. The findings revealed that AI governance and digital leadership positively predicted organizational effectiveness, explaining 38.2% and 45.0% of its variance, respectively. When considered together, both predictors accounted for 57.5% of the variance in organizational effectiveness, with digital leadership showing a stronger relative contribution than AI governance. The study highlights the complementary relevance of responsible AI governance and effective digital leadership to organizational effectiveness and recommends strengthening both capabilities through appropriate institutional policies and continuous capacity-building programmes.

Deep learning-based object detection application for medical images using the vision transformer (ViT) method

2025-10-13T23:49:18+07:00

Ischemic stroke is a medical condition that requires rapid and accurate diagnosis because it can cause permanent brain damage. Medical images, such as CT scans, can serve as a source of information to assist healthcare professionals in identifying patients' brain conditions. However, manual image analysis is time-consuming and may be affected by human limitations in image interpretation. This study aims to develop a CT scan image classification application to assist in identifying ischemic stroke using the Vision Transformer (ViT) method. The research stages include data preprocessing, patch embedding, positional encoding, transformer encoder, and classification head to classify images into two categories: stroke and normal. The dataset consists of patient CT scan images obtained from Hendra Hospital, Binjai. The system was implemented using the Python programming language and the FastAPI framework. The evaluation results show that the ViT model achieved an accuracy of 82.93% in the final testing. These results indicate that the ViT model can be used to classify CT scan images into stroke and normal categories on the dataset used in this study. This research may serve as an approach for developing deep learning-based decision support systems to assist in the analysis of medical images.

A comparative study of GCN and GraphSAGE algorithms for structural anomaly detection in public procurement networks

2026-09-02T21:10:33+07:00

Bid rigging and procurement collusion may produce structural patterns in the relationships between companies and procurement opportunities. This study compares Graph Convolutional Network (GCN) and GraphSAGE for structural anomaly detection in a heterogeneous company–tender procurement network. The study uses Government Procurement via GeBIZ data from Singapore and introduces controlled synthetic anomalies because the dataset does not provide complete bidder-participation information or verified collusion labels. A leakage-aware temporal evaluation is applied together with multiple anomaly difficulties, random seeds, unseen anomaly subjects, and cross-pattern testing. The results show different empirical behavior between the two architectures. GraphSAGE achieves higher F1-Score, ROC-AUC, and PR-AUC in the multi-seed evaluation, while GCN achieves higher recall. GraphSAGE also performs better in the true inductive evaluation on unseen anomaly subjects. An Early Warning System based on GraphSAGE is further developed to rank companies according to their structural anomaly scores. The findings indicate that graph-based models can support structural anomaly screening, while synthetic anomaly results should not be interpreted as evidence of confirmed procurement collusion.