JURNAL FASILKOM

Analisis Segmentasi Perilaku Pembeli Marketplace Shopee Menggunakan Model RFM dan Algoritma K-Means Clustering pada Toko Retail Online XYZ di Yogyakarta

Rezky Amymaztura — 2026-08-30

Perkembangan Marketplace Shopee di Indonesia membuat salah satu Toko Retail Online XYZ yang bergerak di bidang fashion di Yogakarta memerlukan strategi pemasaran yang efektif agar tepat sasaran. Data transaksi shopee toko tersebut hanya berakhir menjadi arsip penjualan dan belum diolah secara optimal sehingga mengalami kesulitan dalam mengelompokkan dan memahami perilaku pelanggan. Penelitian ditujukan untuk melakukan segmentasi perilaku pelanggan menggunakan pendekatan Customer Relationship Management (CRM) analitik dengan model RFM (Recency, Frequency, Monetary) dan Algoritma K-Means Clustering. Penelitian menggunakan 85.781 data transaksi periode 2025, yang kemudian dilakukan tahap preprocessing dan pembersihan outlier dengan Interquatile Range (IQR) sehingga menjadi 43.036 data. Evaluasi gabungan dilakukan untuk menentukan klaster optimal menggunakan metode Elbow, Silhouette Score dan Davies-Bouldin Index. Penelitian menetapkan hasil k=3 sebagai jumlah klaster optimal dengan titik siku pada grafik yang ditunjukkan pada grafik Elbow serta divalidasi menggunakan Silhouette Score dengan nilai paling tinggi 0,4949 dan nilai minimum Davies-Bouldin Index 0,7429. Mayoritas pelanggan memiliki karakteristik one-time buyers, didasari oleh nilai Frequency yang relatif sama sehingga nilai variabel Recency dan Monetary lebih mempengaruhi segmentasi. Segmen yang dihasilkan adalah Klaster 0 (New Customer / Potential), Klaster 1 (At Risk Customer), Klaster 2 (Lost Customer). Penelitian berhasil mengubah ribuan data transaksi menjadi informasi yang strategis dengan menggambarkan karakteristik setiap segmen pelanggan sehingga perusahaan dapat menerapkan strategi pemasaran dengan tepat.

Pengembangan E-Registcam Dengan Penerapan Preprocessing Imagick Untuk Pengenalan Citra KTP Berbasis OCR

Renaldi Riyandi — 2026-08-30

Identity card data entry is commonly performed manually and may result in human errors during the registration process. Optical Character Recognition (OCR) technology can be utilized to automate the extraction of textual information from identity card images. This study aims to develop the E-Registcam system by applying Imagick-based image preprocessing for OCR-based Indonesian identity card (KTP) recognition. The preprocessing stages consist of grayscale conversion, resizing, contrast enhancement, noise removal, sharpening, and cropping. OCR processing is performed using Tesseract OCR to extract the name field from KTP images. The study utilized 30 KTP image samples and evaluated recognition performance using the Character Accuracy Rate (CAR) method. Experimental results showed that OCR with preprocessing successfully recognized 23 images and failed on 7 images, achieving an average CAR value of 74.13%. Meanwhile, OCR without preprocessing successfully recognized 26 images and failed on 4 images, with an average CAR value of 81.05%. The results indicate that the proposed system is capable of extracting textual information from KTP images; however, the applied preprocessing stages did not consistently improve OCR recognition accuracy. Factors such as lighting conditions, image quality, and background complexity affected the recognition performance.

Evaluasi Kritis Random Forest dalam Prediksi Kegagalan CI/CD Pipeline pada Dataset Log Sintetis

Rizqia Fauziah Rachma — 2026-08-30

Continuous Integration and Continuous Deployment (CI/CD) pipelines are crucial in modern DevOps environments; however, pipeline failures often hinder software delivery. This study aims to implement Machine Learning specifically the Random Forest algorithm to predict the specific stage of failure (Build, Test, or Deploy phases) using execution logs. The research methodology encompasses data preprocessing, the extraction of three key dynamic features (execution duration, CPU usage, and memory consumption), model training using an 80:20 train-test split, and performance evaluation via a confusion matrix. Evaluation based on 9,000 test samples representing 20% of the total 45,000 synthetic failure logs yielded an overall accuracy of 33.22%, with F1-scores of 0.33 for the Build class, 0.35 for Deploy, and 0.32 for Test. These results indicate that the Random Forest model possesses very limited generalization capability because the synthetic dataset lacked naturally robust failure trace patterns, as evidenced by an accuracy level approaching that of random guessing. It is concluded that the use of a synthetic log dataset constitutes the primary limitation and the root cause of the model's generalization failure in this study. Therefore, future research is recommended to utilize actual operational logs from real-world CI/CD platforms, such as GitHub Actions or Jenkins, to enable the model to demonstrate more valid causal relationships.

Perbandingan Algoritma Support Vector Machine dan Random Forest untuk Analisis Sentimen terkait Makan Bergizi Gratis (MBG)

Alfian Saputra — 2026-08-30

Advances in digital technology have made social media the primary platform for the public to voice their opinions on government policies, including the Free Nutritious Meals (MBG) program. This study aims to compare the performance of the Support Vector Machine (SVM) and Random Forest algorithms in conducting sentiment analysis of public opinion on the X platform regarding this policy, in order to provide an objective overview for policymakers. A total of 2,164 tweets were collected via crawling using Tweet Harvest from May to August 2025. The methodological steps included preprocessing, BERT-based automatic labeling, Eliminate the neutral class to focus the analysis on binary classification (positive and negative), TF-IDF feature extraction, and the application of SMOTE to address class imbalance in the dataset. Model optimization was performed using Grid Search with a 5-Fold Cross Validation testing scheme and an 80:20 data split. The results of the study indicate that the majority of public responses to the MBG program were positive. Based on the final evaluation, the SVM with a linear kernel proved superior with an accuracy of 78.24% and a macro average F1-score of 0.76, outperforming the Random Forest (300 estimators), which achieved an accuracy of 75.00% and an F1-score of 0.73. The application of SMOTE has proven to be crucial in improving the negative class recall, enabling the model to identify minority sentiments more accurately. This study concludes that SVMs demonstrate more stable and objective generalization capabilities in mapping public opinion following data balancing.

Penerapan Metode SAW untuk Pemilihan Bank Terbaik Berdasarkan Kriteria Layanan Perbankan bagi Mahasiswa UPN “Veteran” Jakarta

Raisadevi Ayunda Putri — 2026-08-30

Chossing appropriate banking services has become increasingly complex for university students due to the swift expansion of Indonesia's digital financial technology. Previous Decision Support System (DSS) studies in banking, such as those relying on the Analytic Hierarchy Process (AHP), exhibit a critical research gap: they predominantly analyze conventional and digital banks in isolation, use restricted criteria, and neglect the student demographic. To address this gap and establish a distinct novelty, this study introduces an integrated DSS model using the Simple Additive Weighting (SAW) method. Unlike existing AHP-based models that rely heavily on subjective expert judgments, this SAW framework contributes uniquely by combining subjective student preferences with objective institutional metrics, including monthly fees, ATM networks, and branch distributions extracted from official annual reports and corporate websites. The model evaluates both banking sectors across six dimensions tailored for students at UPN "Veteran" Jakarta. Prior to SAW processing, the survey instrument spanning seven faculties was validated via Pearson Correlation and Cronbach's Alpha. The findings reveal that SeaBank achieved the highest preference score (0.878), outperforming BCA (0.794) and BRI (0.786). This demonstrates that while students prioritize administrative cost efficiency and digital agility, security and physical accessibility remain critical. Ultimately, this research advances banking DSS literature by providing a highly objective, student-centric hybrid evaluation system.

The Implementasi Paralel K-Means Clustering dan Decision Tree untuk Deteksi Risiko Kredit Macet

Halili Maar — 2026-08-30

Non-Performing Loan (NPL) pose a major threat to the stability of financial institutions. This study implements parallel K-Means Clustering and Decision Tree methods in RapidMiner 2026.0.5 to identify potential credit default risks using the Credit Risk Dataset. The dataset consists of 32,576 borrower records with 12 attributes after filtering. Preprocessing includes handling missing values, dummy coding, converting loan_status into a binary variable, and applying Z-Transformation normalization. The Multiply operator enables K-Means (k=3) and Decision Tree to run simultaneously. The Clustering results show that Cluster 0 (33.10%) is a low-risk group dominated by homeowners with mortgages and an A loan rating; Cluster 1 (32.07%) is a medium-risk group dominated by tenant borrowers with a loan rating of B, and Cluster 2 (34.83%) is a high-risk group dominated by a history of delinquency, loan ratings of C–G, and the highest interest rates. The Decision Tree model achieved an Accuracy of 89.26%, Precision of 76.30%, Recall of 73.05%, and an F-Measure of 74.58%, making it effective as an early detection system for nonperforming loans.

Deteksi Kantuk Mahasiswa pada Pembelajaran Online Menggunakan YOLOv8 dengan Pendekatan CRISP-DM pada Dataset Roboflow

Dennis Ma'rifatul Kurnia — 2026-08-30

The quality of online learning in higher education is highly influenced by the level of student alertness and concentration during the lecture process. However, the phenomenon of drowsiness resulting from physical fatigue and prolonged screen time often reduces learning effectiveness. The limitations of lecturers in visually monitoring students' conditions on online learning platforms drive the need for an automated monitoring system capable of operating in real-time. This study aims to develop a student drowsiness detection system using the YOLOv8 algorithm with a CRISP-DM approach. The dataset used originates from two public datasets on the Roboflow platform, namely SleepyDetect and Yawn, which then underwent curation, labeling, and augmentation processes to produce 8,573 images across three classes: open-eyes, pre-drowsy (yawning), and drowsy (eyes closed). The evaluation results show that the model achieves an mAP50 value of 0.992, Precision of 0.973, Recall of 0.979, and an F1-Score of 0.976, with an inference speed of 4.1 ms per image, demonstrating high detection performance on the utilized dataset. The model's implementation was carried out through the AwakeLens software integrated with a monitoring dashboard to support the online learning process. Nevertheless, because this research relies on a public dataset, the obtained results still require further validation using a primary dataset collected from Indonesian students to enhance the external validity and generalization capability of the model.

Identifikasi Tanaman Obat Indonesia dengan Vision Transformer dan Augmentasi Adaptif

Novi Tristanti — 2026-08-30

Manual identification of medicinal plants faces serious challenges due to morphological similarities between species, variations in lighting, and limited availability of botanists in the field. This study proposes an identification system for 100 types of Indonesian medicinal plants using the Vision Transformer (ViT) architecture with a stepwise fine-tuning approach and adaptive image augmentation. The model used is vit_base_patch16_224 initialized with ImageNet-1k pretrained weights, equipped with a classifier head consisting of a series of LayerNorm, Dropout (0.3), and Linear (768→100). The training strategy integrates stepwise freezing (freeze-unfreeze) in the first three epochs, the AdamW optimizer with a weight decay of 0.05, label smoothing (ε=0.1), and cosine-based learning rate scheduling to ensure stable convergence on medium-scale datasets. The dataset used consists of 10,000 images divided using a non-stratified random split with a ratio of 70% training data, 15% validation data, and 15% test data. The evaluation results show that the model achieved an accuracy of 97.3% on the test data, with a macro precision of 0.974, a macro recall of 0.975, and a macro F1-score of 0.973. The macro values were calculated by summing the metric values for each class separately, then dividing by the number of classes without considering the number of samples in each class. The training process lasted for 16 epochs before being terminated by the early stopping mechanism with the best validation accuracy of 97.53% at the 11th epoch. These results demonstrate that the ViT stepwise fine-tuning approach is able to address the challenges of multi-class scale classification on medium-sized datasets effectively and reproducibly.

Implementasi Metode K-Means Clustering dan Association Rules Apriori untuk Pengelolaan Stok Obat

Nursyamsiah Fajriah — 2026-08-30

Drug inventory management at Klinik Pratama Kencana was previously performed through manual or semi-manual recording, making stock monitoring and procurement decisions less effective. This study develops a web-based drug inventory management system by integrating K-Means Clustering and Apriori Association Rules. Historical stock and drug-out transaction data from January to March 2026 were processed using a descriptive quantitative applied-research approach. K-Means used initial stock, incoming stock, outgoing stock, and remaining stock attributes after Min-Max normalization to classify 64 drugs into fast-moving, medium-moving, and slow-moving groups. Apriori analyzed 597 transaction baskets with a minimum support of 5% and a minimum confidence of 40%. The clustering produced 15 fast-moving, 27 medium-moving, and 22 slow-moving drugs. Apriori generated several strong rules; the rule Acetylcysteine to Ambroxol achieved 11.06% support, 85.71% confidence, and a lift ratio of 5.39. System calculations matched Microsoft Excel validation results, all tested functions were valid, and five users gave a 96% acceptance score. The integrated system provides stock status, restock recommendations, monthly reports, and co-occurrence patterns to support more structured inventory decisions. By combining stock movement classification with drug co-occurrence patterns, the system helps administrators prioritize restocking together with related medicines, thereby reducing manual checking and improving inventory monitoring efficiency.

Prediksi Harga Bitcoin Menggunakan Support Vector Regression (SVR) Berbasis Particle Swarm Optimization (PSO)

Rizky Parlika — 2026-08-30

Bitcoin prices exhibit high volatility and complex non-linear patterns, making accurate price predictions challenging in the context of cryptocurrency market research. In this study, an optimized SVR-PSO predictive modeling approach is developed to help predict Bitcoin prices, and its performance is compared against a Grid Search-optimized SVR baseline. The SVR-PSO modeling approach developed in this study uses daily historical Bitcoin records obtained from Yahoo Finance between March 10, 2019, and May 11, 2026. The data consists of open, high, low, close, and volume values as attributes, and the closed value as the target. Data preprocessing consists of chronologically organizing the data, removing blank rows, min-max scaling, and splitting the data into 80:20 for training and testing. PSO is used to identify the optimal SVR parameters C, ε, and γ with the aim of reducing MAPE. The optimal parameters for the SVR model are C = 501, ε = 0.001, and γ = 0.001. On the test dataset, the SVR-PSO model achieved an MAE of 1,521.47, an RMSE of 2,101.70, a MAPE of 1.65%, and an R² of 0.9813, outperforming the SVR-GridSearch baseline (MAE 1,559.87; RMSE 2,128.52; MAPE 1.69%; R² 0.9809). These results suggest that SVR-based Bitcoin price prediction is more accurate when optimized using PSO than using a discrete Grid Search.

Perbandingan Model Deep Learning LSTM, GRU, dan Bi-LSTM untuk Prediksi Hujan Harian Australia Menggunakan Teknik SMOTE

Ari Risnanto — 2026-08-30

Climate change causing increasingly erratic rainfall patterns, triggering an increase in hydrometeorological disasters such as floods, droughts, and declining agricultural productivity. Therefore, accurate rainfall prediction is crucial for mitigation and decision-making. However, previous research often focuses solely on accuracy metrics without evaluating the model's computational burden, and often ignores the problem of class imbalance in weather datasets. This study evaluates the performance and computational efficiency of LSTM, GRU, and Bi-LSTM deep learning models for daily rainfall prediction using the historical Australian meteorological dataset weatherAus. The novelty of this study lies in the comprehensive mapping between predictive quality and resource efficiency after dataset balancing. The preprocessing stage includes handling missing values, categorical data transformation, data leakage prevention, data sharing, and the application of SMOTE oversampling. The results of the area under the curve (AUC-ROC) evaluation show that the GRU model is superior with a value of 0.85, surpassing LSTM and Bi-LSTM, respectively, at 0.84. In the rain class recall metric, GRU again leads (0.70), compared to LSTM (0.67), and Bi-LSTM (0.57). Computational evaluation, GRU is significantly more efficient with the fastest training time (1,306.26 seconds), followed by LSTM (2,259.12 seconds), and Bi-LSTM (13,348.43 seconds). Peak RAM usage relatively comparable, GRU (2,053.77 MB), LSTM (1,971.47 MB), and the highest Bi-LSTM (2,242.60 MB). These findings conclude that GRU is recommended as the most optimal model that balances accuracy and efficiency, LSTM as an alternative, while Bi-LSTM is considered less effective. Future research recommended to explore hybrid architectures or ensemble learning to capture more complex spatiotemporal patterns.

Pembangunan Multi-Domain Human-Labelled Dataset untuk Konversi Bahasa Alami ke SQL pada Bahasa Indonesia

Fiorella Asyfa Firlanda — 2026-08-30

This study develops a multi-domain, human-labelled Text-to-SQL dataset in Indonesian to address the absence of such resources for the language. The dataset was constructed through a corpus-based approach comprising four stages: corpus planning, corpus construction, corpus validation, and model evaluation. Database schemas represented as Entity Relationship Diagrams (ERD) were used as the structural foundation for generating pairs of Indonesian natural language questions and SELECT-type SQL queries, all created through manual annotation. Validation was performed through four sequential mechanisms: SQL syntax checking, schema conformity verification, query execution testing, and semantic alignment assessment between questions and SQL queries. The resulting dataset is Spider-compatible in JSON format, covering 40 databases, 27 domains, 1,524 questions, and 1,355 unique SQL queries distributed across four difficulty levels. Preliminary evaluation using SQLNet and TypeSQL baseline models under example split and database split scenarios confirms that the dataset provides a representative and challenging evaluation environment for Indonesian Text-to-SQL experiments, though model performance remains limited on complex queries and previously unseen database schemas. The dataset is publicly available and is intended to support future development and evaluation of Indonesian Text-to-SQL models..

Ablasi Kelompok Fitur Multi-View pada Random Forest untuk Deteksi Intrusi IIoT

Januar Al Amien — 2026-08-30

Internet of Things (IIoT) systems generate heterogeneous multisource telemetry data, including network traffic, host resource usage, and security logs. Intrusion detection studies commonly combine all available feature sources based on the assumption that incorporating more sources (multi-view) will always improve detection performance. This study examines this assumption using the X-IIoTID dataset through two experiments. First, feature selection based on Random Forest Gini importance was evaluated using five feature sizes (K = 10, 20, 30, 45, and 61), with cross-algorithm robustness assessed using Decision Tree, Logistic Regression, and K-Nearest Neighbors. Second, a systematic ablation study was conducted on seven combinations of three feature groups: Network (N), Host (H), and Log (L), with Timestamp excluded from the Network group to ensure consistent feature treatment. Using 299,999 samples, comprising 239,999 training and 60,000 test samples across 19 attack classes and a normal class, the results show that multiclass performance increased with the number of features, achieving an F1-macro of 0.876 at K = 10 and 0.912 at K = 61. The ablation study showed that the Full MultiView (N+H+L) achieved the best performance (F1-macro = 0.912), followed by N+H (0.905) and N+L (0.879). The Log group alone yielded low performance (0.098) but provided additional value when combined with Network features. These findings demonstrate that the effectiveness of multi-view intrusion detection depends on feature-source combinations rather than merely the number of sources, highlighting the importance of feature-group ablation in designing IIoT intrusion detection systems.

Analisis Arsitektur Smart Water Management Systems Perkotaan melalui Integrative Literature Review

Hafidz Ardianto — 2026-08-30

This study aims to compile an integrative review of Smart Water Management Systems architecture in urban contexts to identify technology integration patterns, dominant methodological approaches, and recurring research gaps. The method used is an Integrative Literature Review of 16 national and international journal articles indexed in Scopus. The review was conducted through systematic selection, structured data extraction, and comparative-thematic synthesis to map architectural components, analytical models, optimization strategies, and evaluation frameworks. Results show that system architecture evolves in a layered structure integrating IoT for data acquisition, wireless communication networks for connectivity, cloud-edge computing for data processing, and AI/ML for predictive analytics and automated decision-making. Optimization strategies focus on water distribution efficiency, improved operational reliability, and resource loss reduction. The literature also reveals gaps in system interoperability, data standardization, cybersecurity, AI model transparency, and the lack of comprehensive evaluation dimensions. The study concludes that architectural integration and systemic evaluation are essential for developing adaptive, reliable, and sustainable urban Smart Water Management Systems.

Penerapan Metode TOPSIS dalam Penentuan Kelayakan Penerima Beasiswa KIP

Fitri Handayani — 2026-08-30

The selection of Kartu Indonesia Pintar (KIP) scholarship recipients requires an objective and systematic decision-making process because it involves multiple social and economic criteria. Manual selection may cause inaccurate targeting due to subjective assessment. This study applies the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS) method to determine the eligibility of KIP scholarship recipients at Universitas Muhammadiyah Pringsewu and evaluates the conformity of the ranking results with actual recipient data using Spearman Rank Correlation. The data consist of 940 KIP scholarship applicants that were cleaned, encoded, and transformed into sixteen decision criteria. Entropy weighting was used to obtain objective criterion weights, followed by TOPSIS calculation to produce preference values and rankings. The results show that TOPSIS recommended 524 of 540 actual recipients, with a decision conformity percentage of 97.04%. The Spearman correlation coefficient reached 0.8479 with a very small p-value, indicating a very strong and significant relationship between TOPSIS rankings and actual data. These findings indicate that TOPSIS is reliable for supporting targeted KIP scholarship selection.

Sistem Penggajian Terintegrasi Berbasis API pada SMART UMRI

Heru Prambadi — 2026-08-30

Digital transformation in higher education requires integrated information systems that support operational efficiency, data consistency, and accountable governance. Payroll management is a critical administrative process because it involves employee master data, compensation components, taxation, and social security contributions. At Universitas Muhammadiyah Riau, payroll processing previously relied on separated data flows that increased the risk of duplicate entry, inconsistent employee data, delayed payroll generation, and calculation errors. This study aims to examine whether an API-based integrated payroll system connected to SMART UMRI can resolve these problems, using employee and payroll data sourced from the SMART UMRI database. The system was developed using Rapid Application Development and implemented as a RESTful API architecture with HMAC authentication, HTTPS, IP allowlist, role-based access control, and audit trail. Evaluation covered functional testing, API integration, payroll validation, security assessment, and performance testing with Apache JMeter. The results show successful synchronization of employee master data, payroll calculation accuracy with zero difference against manual validation, and stable API performance with 0% error rate in operational testing, confirming that the problems identified before the integration have been resolved. The system improves payroll transparency through digital payslips and traceable logs. The model is feasible for deployment in higher education institutions.

Rancang Bangun Smart Squeeze Cage Berbasis Internet of Things untuk Monitoring Bobot dan Rekomendasi Pakan Ternak

Ghina Rania — 2026-08-30

The development of precision livestock farming requires robust and automated data collection tools to minimize animal stress and improve farm efficiency. Traditional livestock weighing methods often lack immediate data access and do not support dynamic resource planning. This study designs and implements a Smart Squeeze Cage based on the Internet of Things (IoT) integrated with a Random Forest Regressor algorithm for real-time livestock weight monitoring and feed optimization. The system integrates four load cell sensors connected in parallel, an HX711 amplifier, and an ESP32 microcontroller embedded within a customized Squeeze Cage structure. Weight data is transmitted via wireless protocol to a centralized cloud database using Supabase and PostgreSQL. Based on historical data, the Random Forest model automatically predicts livestock weight trends and calculates daily feed requirements to provide intelligent recommendations. Testing results indicate high sensor precision with an accuracy of 97% (error tolerance of 0,5 kg), data transmission latency of 1.2 seconds, and a successful data delivery rate of 99.1%. This system offers a seamless, non-invasive solution for data-driven livestock management in modern farming environments.

Analisis Komparatif Model ARIMA, LSTM, dan GRU dalam Forecasting Harga Bitcoin Berbasis Streamlit

Ridwana Septian Nurardian — 2026-08-30

Bitcoin is a cryptocurrency asset characterized by high volatility and non-linear movement patterns that trigger sudden market risks. Considering that the majority of previous studies have solely focused on mathematical metric evaluations without providing practical solutions, this study aims to compare the performance of the conventional statistical model Autoregressive Integrated Moving Average (ARIMA) with Deep Learning architectures, namely Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU), and to implement the best-performing model into an interactive dashboard using Streamlit. The dataset used encompasses daily Bitcoin closing prices for the 2020–2026 period. The methodology includes data preprocessing, MinMaxScaler normalization, sequential data generation using a 30-day sliding window technique, chronological data splitting (80:20), and evaluation using MAE, RMSE, MAPE, Accuracy, and R² Score. Experimental results prove that ARIMA fails to adapt to non-stationary data, whereas GRU outperforms LSTM in terms of architectural efficiency and precision. The GRU model achieved the best performance with an MAE of IDR 34,773,098.88, an RMSE of IDR 46,684,832.94, a MAPE of 2.26%, an Accuracy of 97.74%, and an R² Score of 0.9661. The GRU model was then successfully implemented into a Streamlit web dashboard that facilitates the real-time visualization of historical and predicted prices. In conclusion, the GRU architecture is the most effective and efficient approach for Bitcoin price forecasting, successfully bridging the gap between theoretical analysis and the availability of practical applications.

Arsitektur Hibrida Ekstraksi Fitur Lokal Edge Computing: Tinjauan Literatur Sistematis

Wildan Zhilal Manafi — 2026-08-30

The proliferation of Internet of Things (IoT) devices and the demand for real-time processing have positioned edge computing as critical infrastructure for deploying deep learning models near the data source. Hardware constraints, including limited memory, narrow computational budgets, and strict power limits, challenge deployment of large-parameter neural architectures. This review examines hybrid architectures integrating convolutional neural networks (CNN) with attention mechanisms for local feature extraction on resource-constrained edge devices. Following the PRISMA 2020 protocol, a multi-stage search was conducted exclusively on Scopus, yielding 3,571 records, screened until 121 high-quality studies were included. The review addresses five research questions covering architectural trends, efficiency strategies, performance trade-offs, application domains, and federated learning for privacy-preserving deployment. Findings show that hybrid CNN-Attention architectures outperform pure CNN and Transformer baselines, with 3.2% average accuracy improvement while maintaining competitive inference latency on platforms such as NVIDIA Jetson and Raspberry Pi. Depthwise separable convolution, efficient channel attention, and token aggregation emerged as dominant compression strategies, with health imaging and fault diagnosis as leading domains. The review concludes that hybrid architectures represent the state of the art for edge-oriented local feature extraction, with future directions toward heterogeneous node generalization and multi-modal sensor fusion for distributed IoT networks.

Perbandingan Algoritma LSTM dan GRU dalam Prediksi Hasil Panen Padi Pulau Sumatera

Sapto Hadi Riono — 2026-08-30

Rice production is a strategic commodity that plays a vital role in maintaining food security in Indonesia. Rice production data, being time-series in nature, is influenced by trends, seasonal patterns, and weather and climate factors; therefore, a prediction model capable of accurately capturing temporal patterns is required. This study aims to compare the performance of Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models in predicting rice production on the island of Sumatra. The study utilizes secondary data obtained from Statistics Indonesia (BPS), the Meteorology, Climatology, and Geophysics Agency (BMKG), and Kaggle. Preprocessing stages include handling missing data, data normalization, and transformation into a time-series format. The data is then split into training and testing sets using an 80:20 ratio. Model performance is evaluated using Mean Absolute Error (MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). The results indicate that the GRU model outperforms the LSTM model, yielding an MAE of 0.0505, MSE of 0.0075, and RMSE of 0.0868, whereas the LSTM model produces an MAE of 0.0522, MSE of 0.0082, and RMSE of 0.0917. Thus, the GRU model is more effective at capturing the temporal patterns of rice production data and generating more accurate predictions compared to the LSTM model. These findings are expected to serve as a reference for selecting rice production prediction models to support data-driven decision-making in the agricultural sector.

Algoritma K Means dan TF-IDF untuk Pengelompokkan Opini Publik terhadap Program Makan Bergizi Gratis pada Komentar TikTok

Devi Sugianti — 2026-08-30

In Indonesia, 14% of children suffer from stunting due to malnutrition. To address this issue, the government launched the Free Nutritious Meal (MBG) Program, which has generated diverse public opinions on social media, particularly TikTok. This study aims to cluster public opinions regarding the MBG program using the K-Means Clustering algorithm combined with Term Frequency–Inverse Document Frequency (TF-IDF) without requiring manual labeling. A total of 57,362 comments were collected, of which 53,204 valid comments remained after preprocessing. Truncated Singular Value Decomposition (SVD) was applied for dimensionality reduction, while the optimal number of clusters (K = 5) was determined using the Elbow Method and Silhouette Score. The clustering results identified five main discussion themes: general program discussion (28.4%), child nutrition and school access (7.9%), spontaneous reactive responses (45.2%), criticism and rejection of the program (15.0%), and support for public figures (3.6%). The model achieved a Silhouette Score of 0.0385 and a Davies–Bouldin Index of 4.9507, reflecting the characteristics of short and informal social media text. The findings demonstrate that the unsupervised clustering approach effectively maps public opinion into meaningful thematic groups and provides valuable insights for the National Nutrition Agency to improve menu quality, budget transparency, and distribution standards of the MBG program.

Penerapan TOGAF ADM untuk Pemodelan Blueprint Arsitektur Enterprise pada SMK Berinfrastruktur TI Minim

Wide Mulyana — 2026-08-30

The integration of information and communication technology (ICT) in education is increasingly important, yet many vocational secondary schools (SMK) in Indonesia still lack adequate IT infrastructure and enterprise architecture planning. SMK Sulthan Muazzam Syah Pekanbaru, accredited “C” by the National Accreditation Board (BAN), is one such institution that operates without a systematic IT blueprint to guide technological development. This research aims to design an enterprise architecture-based IT blueprint for the school using The Open Group Architecture Framework Architecture Development Method (TOGAF ADM), covering six phases selected from the full TOGAF ADM cycle: Preliminary, Architecture Vision, Business Architecture, Information Systems Architecture, Technology Architecture, and Opportunities & Solutions. Data were collected through structured interviews with five key stakeholders and direct observation. The results produced four main architectural artifacts: an enterprise architecture blueprint encompassing business, data, application, and technology domains; three entity-relationship diagrams for the online admissions, e-learning, and academic information systems; phased technology infrastructure recommendations, including hardware and network topology upgrades; and a seven-stage, three-month system implementation roadmap. This study contributes an operational enterprise architecture model for resource-constrained vocational schools in Indonesia, addressing a gap in prior TOGAF ADM research that has largely focused on universities and private organizations. The proposed blueprint has the potential to serve as a reference governance framework for similar institutions undergoing ICT transformation, although its transferability to other contexts still needs to be tested empirically.

Prediksi Keberhasilan Akademik Siswa Berbasis Fitur Kategorikal Na-tive dengan Explainable AI (SHAP) menggunakan CatBoost vs LightGBM

Bayu Anugerah Putra — 2026-08-30

Prediction of students academic success is important to support decision-making in education. Educational datasets are generally dominated by categorical variables that require encoding before modeling, which may cause information loss and reduce accuracy. This study applies the CatBoost algorithm, which processes categorical variables natively without additional encoding, to predict students' Exam Score on the Student Performance Factors dataset from Kaggle, with LightGBM used as a comparison model. Evaluation was carried out under three data-split schemes (70:30, 80:20, 90:10) using k-fold cross-validation and three regression metrics (R², MAE, RMSE), followed by model interpretation using Shapley Additive Explanations (SHAP). The results show that CatBoost consistently outperforms LightGBM across all schemes, with the best performance obtained under the 90:10 scheme (CatBoost: R² = 0.851, MAE = 0.475, RMSE = 1.414; LightGBM: R² = 0.809, MAE = 0.758, RMSE = 1.599). SHAP analysis identifies Attendance, Hours_Studied, and Previous_Scores as the most influential features in the prediction. These findings confirm that combining CatBoost with SHAP produces an academic prediction model that is both accurate and transparen.

Multi-Horizon Bitcoin Volatility Forecasting Using Support Vector Machines

Mohammad Faris Al Fatih — 2026-08-31

This study uses Support Vector Regression (SVR) to forecast Bitcoin realized volatility across four timeframes (2h, 4h, 8h, and daily) and four forecast horizons using Binance OHLCV data from January 2023 to June 2026. Realized volatility is estimated using a 5-period rolling standard deviation of log returns. We construct 14 features from lagged volatility, returns, trading volume, RSI, MACD, Bollinger Bandwidth, and high-low spread. We evaluate linear and radial basis function (RBF) kernels using walk-forward validation with a 100-sample rolling window. To prevent look-ahead bias, the training window is adjusted for each forecast horizon so that only targets observable at the forecast time are included. Performance is evaluated using MAE, RMSE, MAPE, and R². The linear kernel outperforms the RBF kernel across all 16 timeframe-horizon combinations, achieving the highest R² of 0.5140 for daily volatility at the shortest horizon. Short horizons () generally produce positive R², whereas longer horizons () yield near-zero or negative R². The look-ahead bias correction produces only modest changes in performance, indicating that the initial bias had limited influence on the results. Permutation importance identifies lagged volatility as the dominant predictor, supporting volatility persistence. Overall, the findings demonstrate that linear SVR provides a simple, effective, and interpretable approach to Bitcoin volatility forecasting.

Pengembangan Model Klasifikasi Code Smells Pada Backend Python Menggunakan Algoritma Random Forest (Studi Kasus Proyek Open Source Github)

Aqilla Ar-Hammar — 2026-08-30

Deadline pressure in software development often drives coding shortcuts, leading to internal quality degradation known as code smells. These structural anomalies contribute to technical debt accumulation and complicate system maintenance over time. This study develops an automated classification model to detect code smell contamination in the Python backend ecosystem. The methodology uses the Random Forest ensemble algorithm integrated with the Synthetic Minority Over-sampling Technique (SMOTE) for class balancing. Data mining on GitHub with high-reputation criteria extracted 137,728 code samples from 7 large-scale repositories using the Radon multi-metric tool. To simulate human error in real-world scenarios, 5% random noise was inserted into the labeling data. Testing using the confusion matrix shows the proposed model achieves highly stable and balanced performance, with average precision, recall, and f1-score of 0.95 in both macro and weighted averages. Ablation study analysis proves that SMOTE intervention effectively maintains detection consistency in minority class categories. Feature importance ranking identifies the Logical Lines of Code (LLOC) metric as the most crucial indicator with 37.65% influence weight, followed by LOC and Blank metrics. This research provides an automated quality assurance system for developers to detect code refactoring opportunities at an early stage.