Garuda - Garba Rujukan Digital

Prediction of Export Volume in South Sulawesi Based on Destination Country Using the BPNN Method Widiyanti, Widiyanti; Anggreani, Desi; Lukman, Lukman; Danuputri, Chyquitha
Jurnal Algoritma, Logika dan Komputasi Vol 8, No 2 (2025)
Publisher : Universitas Bunda Mulia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.30813/j-alu.v8i2.8782

This study aims to develop a prediction model for South Sulawesi's export volume based on destination countries using the Backpropagation Neural Network (BPNN) method. South Sulawesi has a significant contribution in the export of agricultural, marine, and mining commodities to various Asian and global countries. Common problems in the export process are unpreparedness of goods, limited commodity stocks, and a mismatch between production capacity and destination market demand. The export data used is from 2018 to 2026 with a total of 1,555 rows of data . The BPNN model with a 6-6-1 architecture is applied to study historical patterns and make accurate predictions. The test results show a Mean Squared Error (MSE) value of 0.0161440, with prediction results close to the actual trend. Exports peaked at almost 40 tons in 2019 and decreased significantly in 2023, then are predicted to recover steadily in 2025–2026. The destination countries with the highest export volumes are China, Japan, and East and Southeast Asian countries. The main commodities contributing significantly are octopus, processed wood, and marine products. These findings demonstrate that the BPNN method is effective in identifying export patterns and can be used as a basis for data-driven trade planning at the regional level. This study also underscores the importance of logistical readiness and market diversification in efforts to maintain export sustainability.

Model Deep Learning Berbasis Convolutional Neural Network Untuk Identifikasi Stroke Iskemik Pada Citra CT Scan Faturohman, Agung; Anggreani, Desi; Yusliana Bakt, Rizki
Jurnal Pustaka AI (Pusat Akses Kajian Teknologi Artificial Intelligence) Vol 5 No 2 (2025): Pustaka AI (Pusat Akses Kajian Teknologi Artificial Intelligence)
Publisher : Pustaka Galeri Mandiri

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.55382/jurnalpustakaai.v5i2.1150

Stroke iskemik merupakan salah satu penyakit tidak menular yang berbahaya dan dapat menyebabkan kecacatan hingga kematian apabila tidak ditangani dengan cepat dan tepat. Identifikasi stroke melalui citra CT scan otak menjadi metode penting dalam dunia medis, namun masih memerlukan waktu dan keahlian tinggi. Penelitian ini bertujuan untuk mengembangkan sistem deteksi stroke iskemik secara otomatis menggunakan metode Convolutional Neural Network (CNN) dengan arsitektur MobileNetV2. Data yang digunakan berupa citra CT scan otak pasien dari Rumah Sakit Labuang Baji Makassar, yang diproses melalui tahapan preprocessing seperti grayscale, resizing, augmentasi, dan normalisasi. Model CNN dilatih menggunakan binary crossentropy loss dan Adam optimizer untuk klasifikasi dua kelas, yaitu normal dan stroke iskemik. Hasil pengujian menunjukkan bahwa model mencapai akurasi sebesar 91,6%, precision 88%, recall 95,1%, dan F1-score 0,914, yang menandakan bahwa model ini mampu mengenali stroke iskemik secara efektif. Dengan demikian, sistem ini berpotensi menjadi alat bantu diagnosis awal yang efisien dan akurat dalam bidang kesehatan.

Stacking architecture-endpoint detection: a hybrid multi layered architecture for endpoint threat detection Wahid, Abd Rahman; Anggreani, Desi; Hayat, Muhyiddin A. M.; Abd Rahman, Aedah; Faisal, Muhammad
International Journal of Advances in Applied Sciences Vol 14, No 4: December 2025
Publisher : Institute of Advanced Engineering and Science

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.11591/ijaas.v14.i4.pp1263-1280

Modern endpoint threat detection systems face persistent challenges in balancing detection accuracy, resilience against zero-day attacks, and the interpretability of artificial intelligence (AI) models. Although deep learning (DL) approaches often achieve high accuracy on benchmark datasets, they remain vulnerable to adversarial perturbations and operate as opaque “black boxes,” thereby reducing trust and limiting practical adoption in critical infrastructures. This research introduces stacking architecture-endpoint detection (STACK-ED), a hybrid multi-layered architecture for endpoint threat detection. STACK-ED integrates three complementary paradigms: supervised learning for known attack patterns, self-supervised Fgraph-based learning for structural relationships, and unsupervised anomaly detection for emerging or unknown threats. The outputs are consolidated by a meta learner, followed by a post-hoc correction (PHC) mechanism to minimize false negatives. The framework was evaluated on a combined benchmark dataset (CSE-CIC-IDS2018 and UNSW-NB15, hereafter referred to as HIDS-Set). Experimental results demonstrate state-of-the-art performance, achieving an F2-score of 98.89% after hybrid integration and active learning, with the primary optimization objective being the reduction of undetected attacks. Furthermore, the Shapley additive explanations (SHAP) method enhances interpretability by revealing feature contributions, while the PHC successfully recovered 62.64% of missed zero-day candidates. The findings position STACK-ED not only as a highly accurate detection model but also as an adaptive, resilient, and transparent framework, offering practical implications for enterprise-grade endpoint defense and future zero-trust cybersecurity systems.

A Hybrid Convolutional Neural Network and Bidirectional LSTM Architecture for Multi-Sector Export Forecasting: A Macroeconomic Time Series Analysis of Indonesia Desi Anggreani; Nurmisba Nurmisba; Aedah Abd Rahman
Indonesian Journal of Data and Science Vol. 6 No. 3 (2025): Indonesian Journal of Data and Science
Publisher : yocto brain

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.56705/ijodas.v6i3.330

Accurately predicting export values is key for a country in formulating its economic plans. Unfortunately, export data often exhibits complex time series patterns that are difficult to predict, characterized by non-linearity, high volatility, and complex temporal dependencies. This study offers a solution by testing a combined deep learning model, specifically a fusion of Convolutional Neural Networks (CNN) and Bidirectional Long Short-Term Memory (BiLSTM), to address the challenges of export time series forecasting. This study uses this approach to forecast Indonesia's monthly export time series data from 2016 to 2023, covering various sectors ranging from oil and gas, non-oil and gas, agriculture, industry, mining, and others. The core idea is to leverage the CNN's ability to identify hidden features within time series patterns, while the BiLSTM is tasked with understanding the temporal flow of data from both directions to capture the inherent long-term temporal dependencies within economic time series data. As a result, this combined model proved to be far superior to the standard BiLSTM model in handling the complexity of export time series. In the Non-Oil and Gas sector, the proposed model achieved a high level of accuracy with an MSE value of 3,330,239.74, an RMSE of 1,824.89, and an average prediction error (MAPE) of only 8.17%, representing a significant improvement of 69% over the baseline BiLSTM model. Similar success was also found in all other sectors, proving that this hybrid approach is highly promising for complex economic time series analysis

Fine-Tuning a Large Language Model on Vertex AI for a New Student Registration Chatbot at Universitas Muhammadiyah Makassar Desi Anggreani; Muhyiddin A M Hayat; Lukman; Ahmad Faisal; Khadijah; Darniati
Indonesian Journal of Data and Science Vol. 7 No. 1 (2026): Indonesian Journal of Data and Science
Publisher : yocto brain

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.56705/ijodas.v7i1.341

This study addresses the limitations of manual admission services at Universitas Muhammadiyah Makassar, which often result in delayed and inconsistent information delivery. To overcome these challenges, an institution-specific chatbot was developed by fine-tuning the Gemini 2.5 Flash model on the Google Cloud Vertex AI platform. The model was trained using a curated domain-specific dataset of 1,430 question–answer pairs derived from official documents and frequently asked questions. The fine-tuning process employed supervised learning to enhance contextual relevance and response accuracy. System performance was evaluated using automated text quality metrics, achieving an average BLEU score of 0.23526 and a ROUGE-L Recall score of 0.53424, indicating satisfactory lexical and semantic similarity. Furthermore, a user acceptance evaluation involving 52 respondents yielded a Customer Satisfaction Score (CSAT) of 84.2%, reflecting high user satisfaction. These results demonstrate that fine-tuning a Large Language Model (LLM) for specific institutional needs effectively improves both response quality and service reliability. Ultimately, this approach offers a practical and scalable solution for modernizing student admission services in higher education, ensuring that prospective students receive accurate information in a timely and efficient manner.

A Hybrid BERT–RAG Model for Developing Knowledge-Validated Conversational Systems Anggreani, Desi; Ismawati, Ismawati; Auliyah, A. Inayah; Lukman, Lukman; Rahman, Aedah Abd; Nurmisba, Nurmisba; Akbar, Muh Ilham
ILKOM Jurnal Ilmiah Vol 18, No 1 (2026)
Publisher : Prodi Teknik Informatika FIK Universitas Muslim Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.33096/ilkom.v18i1.3126.30-42

The transition of freshmen into the university environment requires adaptive and responsive information support. This study develops a chatbot system based on a hybrid BERT–RAG architecture integrated with the FAISS Index to provide automated consultation services for new students. The novelty of this research lies in the implementation of a faculty-based hierarchical knowledge structure and an adaptive multi-domain context mechanism—an approach not previously found in studies involving BERT–RAG for university onboarding services. This design enables the chatbot to deliver more relevant, personalized, and faculty-specific responses. The dataset was derived from three primary sources of information: the Faculty of Economics and Business (FEB), the Faculty of Teacher Training and Education (FKIP), and the Faculty of Engineering (FT), which were structured into a validated knowledge base in documents.json format. System evaluation was conducted across ten interaction scenarios using performance metrics including BERT Similarity, BLEU Score, ROUGE-1, ROUGE-2, and ROUGE-L. The system achieved excellent results, with average scores of 0.905 (BERT Similarity), 0.844 (BLEU), 0.876 (ROUGE-1), 0.820 (ROUGE-2), and 0.871 (ROUGE-L) and standard deviations below 0.1 across all metrics. Strong metric correlations (0.85–0.99) further indicate consistency between semantic understanding and generated text quality. Furthermore, the system effectively minimizes hallucination through validated knowledge integration and faculty-based reranking strategies. Overall, this research provides a significant contribution to the development of institutionally contextual educational chatbots capable of delivering accurate, natural, and responsive communication to support new student orientation in higher education

KLASIFIKASI TANAMAN OBAT TRADISIONAL BERBASIS CITRA BUAH DAN DAUN Kusumawardani, Nurul; Danuputri, Chyquitha; Darniati; Faisal, Muhammad; A.M Hayat, Muhyiddin; S. Kuba, Muhammad Syafaat; Anggreani, Desi
PROGRESS Vol 18 No 1 (2026): April
Publisher : P3M STMIK Profesional Makassar

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.56708/progres.v18i1.534

Indonesia is a megabiodiversity country with extensive use of traditional medicinal plants; however, plant identification in natural environments remains largely manual and error-prone. Recent advances in deep learning, particularly Vision Transformer (ViT), provide a promising solution by effectively capturing global spatial features for image classification. This study applies a ViT-Base/16 model to automatically classify fruit and leaf images of Indonesian medicinal plants. The dataset comprises 1,000 field-collected images from Galung Village, West Sulawesi, covering 20 classes (10 medicinal and 10 non-medicinal plants). The model was fine-tuned using the AdamW optimizer with a learning rate of 2×10⁻⁵ and trained for 30 epochs with cosine annealing. The proposed approach achieved high performance, with 99.33% accuracy, 99.41% precision, 99.33% recall, and a 99.33% F1-score, while binary classification between medicinal and non-medicinal plants reached 100% accuracy. The system was deployed as a Flask-based web application, demonstrating reliable functionality and practical response times. Overall, the results confirm the effectiveness of Vision Transformer for medicinal plant classification under natural conditions and highlight its potential to support digital documentation, education, and the preservation of local ethnobotanical knowledge.

Identifikasi Penyakit Tuberkulosis pada Citra X-Ray Paru Menggunakan Swin Transformer dengan Pendekatan Out-of-Distribution Detection Andi Citra Ayu Lestari; Desi Anggreani; Muhammad Faisal; Chyquitha Danuputri
Arus Jurnal Sains dan Teknologi Vol 4 No 1: April (2026)
Publisher : Arden Jaya Publisher

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.57250/ajst.v4i1.2731

Penelitian ini bertujuan mengimplementasikan arsitektur Swin Transformer untuk mengidentifikasi penyakit tuberkulosis pada citra X-Ray paru serta mengevaluasi kemampuan pendekatan Out-of-Distribution Detection dalam mengenali citra yang tidak sesuai dengan distribusi data pelatihan. Data penelitian terdiri atas 1.272 citra yang terbagi seimbang ke dalam tiga kelas, yaitu Tuberkulosis, Non-Tuberkulosis, dan Tidak Dikenali/Out-of-Distribution Detection. Tahapan penelitian meliputi seleksi kualitas citra, resize ke ukuran 384 x 384 piksel, augmentasi data latih, pelatihan model Swin Transformer, serta evaluasi menggunakan accuracy, macro F1-score, confusion matrix, dan AUC. Hasil penelitian menunjukkan bahwa model Swin Transformer tunggal memperoleh accuracy 83,59% dan macro F1-score 83,59% pada klasifikasi Tuberkulosis dan Non-Tuberkulosis. Model Hybrid Swin Transformer + Out-of-Distribution Detection menghasilkan kinerja lebih baik dengan accuracy 89,06%, macro F1-score 89,10%, dan Out-of-Distribution Detection Rate 98,44%. Temuan ini menunjukkan bahwa integrasi Out-of-Distribution Detection mampu meningkatkan keandalan sistem karena model tidak memaksakan prediksi terhadap citra yang tidak relevan. Sistem yang dikembangkan dapat digunakan sebagai alat bantu skrining awal, dengan keputusan akhir tetap berada pada tenaga medis.

A Calibrated ROI-Aware Hybrid CNN-Transformer for Kidney Stone Presence Classification on Heterogeneous Axial CT Images Muh Ilham Akbar; Muhammad Faisal; Desi Anggreani; Abd Rakhim Nanda; Try Gustaf Said; Muhammad Syafaat S. Kuba
Buletin Sistem Informasi dan Teknologi Islam (BUSITI) Vol 7, No 2 (2026)
Publisher : Universitas Muslim Indonesia

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.33096/busiti.v7i2.3463

Batu ginjal merupakan penyebab umum nyeri pinggang akut, dan CT non-kontras menjadi standar referensi untuk mendeteksi kalkulus. Pada penelitian ini, istilah heterogen merujuk pada variasi protokol akuisisi antarrumah sakit, seperti perbedaan dosis radiasi, ketebalan irisan, rekonstruksi, dan bidang pandang, yang dapat mengubah tampilan citra serta menurunkan konsistensi pembacaan. Penelitian ini mengusulkan model hibrida CNN-Transformer yang sadar ROI (implisit) untuk klasifikasi keberadaan batu ginjal pada citra CT aksial heterogen. Arsitektur menggabungkan EfficientNet-B3, encoder Transformer ringan, dan Convolutional Block Attention Module (CBAM) tanpa anotasi ROI manual. Dataset terdiri dari 3.364 citra (1.577 batu, 1.787 non-batu) dengan pemisahan bertingkat 70/15/15. Evaluasi mencakup akurasi, presisi, sensitivitas, spesifisitas, F1, ROC-AUC, PR-AUC, inspeksi kalibrasi, dan audit Grad-CAM. Hasil menunjukkan bahwa penambahan Transformer meningkatkan kinerja dibanding baseline CNN, sedangkan CBAM menggeser profil kesalahan ke sensitivitas yang lebih tinggi. Varian Hybrid+Attention mencapai akurasi 0,9861, F1 0,9851, dan ROC-AUC 0,9967 pada set uji, dengan jumlah negatif palsu lebih rendah dibanding varian hibrida tanpa perhatian. Temuan ini menunjukkan potensi model sebagai alat bantu dokter untuk triase dan pembacaan awal yang lebih konsisten pada data lintas protokol, meskipun validasi eksternal, pemisahan berbasis pasien, dan metrik kalibrasi kuantitatif masih diperlukan sebelum klaim kesiapan klinis.

KLASIFIKASI TANAMAN OBAT TRADISIONAL BERBASIS CITRA BUAH DAN DAUN Nurul Kusumawardani; Chyquitha Danuputri; Darniati; Muhammad Faisal; Muhyiddin A.M Hayat; Muhammad Syafaat S.Kuba; Desi Anggreani
PROGRESS Vol 18 No 1 (2026): April
Publisher : P3M STMIK Profesional Makassar

Show Abstract | Download Original | Original Source | Check in Google Scholar

Indonesia is a megabiodiversity country with extensive use of traditional medicinal plants; however, plant identification in natural environments remains largely manual and error-prone. Recent advances in deep learning, particularly Vision Transformer (ViT), provide a promising solution by effectively capturing global spatial features for image classification. This study applies a ViT-Base/16 model to automatically classify fruit and leaf images of Indonesian medicinal plants. The dataset comprises 1,000 field-collected images from Galung Village, West Sulawesi, covering 20 classes (10 medicinal and 10 non-medicinal plants). The model was fine-tuned using the AdamW optimizer with a learning rate of 2×10⁻⁵ and trained for 30 epochs with cosine annealing. The proposed approach achieved high performance, with 99.33% accuracy, 99.41% precision, 99.33% recall, and a 99.33% F1-score, while binary classification between medicinal and non-medicinal plants reached 100% accuracy. The system was deployed as a Flask-based web application, demonstrating reliable functionality and practical response times. Overall, the results confirm the effectiveness of Vision Transformer for medicinal plant classification under natural conditions and highlight its potential to support digital documentation, education, and the preservation of local ethnobotanical knowledge.

Title

Found 22 Documents
Search

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Title Search

Found 22 Documents Search

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Abstract

Title

Found 22 Documents
Search