Jurnal Riset Informatika

COMPARISON OF PERFORMANCE AND COMPUTATIONAL COMPLEXITY OF CNN AND RESNET50 FOR PNEUMONIA CLASSIFICATION

2026-09-22T07:43:09+00:00

Pneumonia is an acute infection of the lung tissue and the leading cause of death among children under five years of age worldwide. Diagnosis based on chest X-ray images is considered prone to misinterpretation, particularly in mild cases that appear similar to normal lung conditions. CNN is one of the most widely used deep learning models in medical image analysis; however, selecting the appropriate architecture between scratch-built models and transfer learning requires further evaluation, particularly regarding the trade-off between classification performance and computational complexity, an aspect that remains largely unaddressed in prior studies on medical image classification. This study aims to compare the classification performance and computational complexity of a CNN and a pre-trained ResNet50 model using transfer learning for binary pneumonia classification on chest X-ray images. The dataset used is the Chest X-Ray Images (Pneumonia) from Kaggle (Mooney, 2018), comprising 5,856 images: 1,583 in the NORMAL class and 4,273 in the PNEUMONIA class. Both models were trained under the same conditions and evaluated using the metrics accuracy, precision, recall, F1-score, ROC-AUC, and computational complexity, measured by model size, number of parameters, training time, and memory consumption. The results show that the CNN outperforms the ResNet50 across all classification metrics, achieving 98,81% accuracy, 99,07% precision, 99,30% recall, 99,18% F1-score, and 0,9978 AUC. In terms of computational complexity, the CNN has a smaller model size, shorter training time, and lower memory consumption

IMPLEMENTATION OF SOFTWARE DEFINED NETWORKING FOR SIMULATION AND ANALYSIS OF DDOS ATTACKS ON SCHOOL NETWORKS

2026-09-23T07:43:14+00:00

The development of digital services in school environments has increased the importance of network security, particularly in addressing Distributed Denial of Service (DDoS) attacks that can disrupt network service availability. This study aims to develop a school network security simulation based on Software Defined Networking (SDN) and analyze the characteristics of normal traffic and DDoS attack traffic. The research method employed is an experimental approach through network simulation using the Mininet emulator and Ryu Controller. The study began with identifying school network security requirements, followed by designing the SDN network topology, implementing the simulation environment, generating normal traffic, and simulating DDoS attacks using the TCP SYN Flood method. Data collection was conducted using Wireshark, tcpdump, and SDN controller logs to obtain traffic parameters such as throughput, packet loss, delay, jitter, and packet rate. The results showed that normal traffic had an average throughput of 320 Kbps with stable network conditions, while during DDoS attacks the throughput increased significantly to 9.8 Mbps, accompanied by substantial increases in packet loss, delay, and jitter. These findings indicate that DDoS attacks directly affect the quality of school network services. Furthermore, the SDN architecture provides centralized traffic monitoring capabilities, making it a promising approach for developing DDoS detection and mitigation systems in school networks.

IMPLEMENTATION OF RETRIEVAL-AUGMENTED GENERATION BASED CHATBOT FOR EXPORT COMPANY CUSTOMER SERVICE AUTOMATION

2026-09-18T07:37:18+00:00

This study discusses the issue of customer service in export companies, which is still limited in terms of service availability and the ability to handle recurring inquiries from international customers. Limitations in human resources result in responses to customer inquiries not being performed optimally, especially outside of operating hours. This study aims to design, implement, and evaluate a Retrieval-Augmented Generation (RAG)-based chatbot to support the automation of customer service in export companies using the Botika Agentic Platform. The study contributes by implementing and quantitatively evaluating a RAG-based chatbot on a Platform as a Service (PaaS) environment using the DeepEval framework with the LLM-as-a-Judge approach. The methods used in this study include four stages, namely data collection through interviews and analysis of company documents, system design, implementation through the Botika platform, and testing using the DeepEval framework with the LLM-as-a-Judge approach. The testing dataset consists of 93 questions covering four knowledge base domains, namely company profile, export and shipping procedures, partnership information, and product information. Evaluation results indicate that the chatbot achieved an average Answer Relevancy score of 0.98, Faithfulness of 1.00, and Contextual Relevancy of 0.88. The system performs well, with high Answer Relevancy and Faithfulness scores, while Contextual Relevancy is relatively lower, particularly in the product information domain. This is due to the dense structure of product documents, which contain many product variations within a single document, thereby affecting the accuracy of the retrieval process. Overall, the proposed RAG-based chatbot demonstrates its potential to support customer service automation in export companies by providing accurate and contextually grounded responses, while highlighting the importance of knowledge base organization for improving retrieval performance.

DESIGN OF AN IOT-BASED GREENHOUSE MICROCLIMATE MONITORING SYSTEM WITH REAL-TIME NOTIFICATION AND THRESHOLD ALERT

2026-09-23T07:43:47+00:00

Greenhouse microclimate monitoring is essential to maintain plant productivity, yet it is still commonly performed manually or with poorly integrated systems. This research designs and implements an IoT-based Greenhouse microclimate monitoring system using an ESP32 microcontroller integrated with five sensors (DHT22, BMP280, DS18B20, BH1750, and MQ-135), equipped with an LCD display, MicroSD data storage, a web dashboard, and real-time notifications via Telegram and buzzer as threshold alerts. Testing was carried out through sensor calibration/comparison against standard instruments, functional testing, and a 7-day field test at Hydroponic Greenhouse of KWT Mawar 8, Tangerang City, with most sensors well calibrated against the standard instruments. The field test shows high system reliability, with 98.81% data completeness and no significant downtime, and successfully detected microclimate conditions outside the ideal range, particularly air and water temperature during the day and humidity at night, which could potentially cause heat stress, root rot, and fungal disease risk in lettuce plants. The threshold alert feature via Telegram and buzzer is validated to function properly in providing early warnings under real operating conditions

Keywords: IoT, Greenhouse, microclimate monitoring, ESP32, threshold alert, Telegram

COMPARISON OF MOBILENETV2 AND RESNET50 PERFORMANCE IN CLASSIFICATION OF SHALLOT PLANT DISEASES

2026-09-22T07:43:36+00:00

Red onions (Allium ascalonicum L.) are a high-value agricultural commodity in Indonesia, but their productivity is often compromised by plant diseases, particularly yellow spot and leaf blight. Visual identification in the field is subjective, prone to error, and time-consuming. Using artificial intelligence, this problem is addressed with a Convolutional Neural Network (CNN) trained to identify disease types from images. For this purpose, this study applies transfer learning to two CNN architectures, MobileNetV2 and ResNet50, to classify red onion leaf diseases into three classes yellow spot, leaf curl, and healthy leaves. A dataset of 1,752 field-captured, pre-labeled images sourced from a public Roboflow repository was split into 70% for training, 20% for validation, and 10% for testing. Both models were trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32. Evaluation was conducted using accuracy, precision, recall, and F1-score metrics based on the confusion matrix. The results show that ResNet50 achieved a test accuracy of 93.33% and an average F1-score of 0.93, outperforming MobileNetV2, which achieved 90.26% accuracy and an average F1-score of 0.90. MobileNetV2 excels in computational efficiency, achieving faster training times in field implementation than ResNet50, which required longer training times for the 20 epoch run. However, in general, ResNet50 demonstrates better generalization, with a stable training curve, and outperforms MobileNetV2 in terms of accuracy, precision, recall, and F1-score for the classification of onion diseases based on leaf images.

HYBRID SOBEL-BASED CONVOLUTIONAL NEURAL NETWORK AND LEARNING VECTOR QUANTIZATION FOR CHILI SEED IMAGE CLASSIFICATION

2026-09-22T07:43:22+00:00

This study develops an image classification system for chili pepper seeds using a hybrid approach that combines a manual Convolutional Neural Network (CNN) based on Sobel filters and Learning Vector Quantization (LVQ). The dataset consists of 900 images from three chili pepper varieties, namely gendot, raksasa, and rawit, with 300 images for each class captured using a smartphone camera. Unlike conventional CNN architectures that rely on trainable convolutional kernels, this study adopts predefined Sobel filters as fixed feature extractors combined with Learning Vector Quantization as the classifier, resulting in a lightweight classification framework that does not require a training process in the feature extraction stage. The CNN architecture employs three convolution-pooling layers (depth-3), producing feature maps of size 16×16, which are then transformed into 256-dimensional feature vectors through a flattening process. LVQ training was conducted for 30 epochs with an initial learning rate of 0.001 and five prototypes assigned to each class. The model was evaluated using a 5-fold cross-validation scheme and obtained a mean accuracy of 80.78% with a standard deviation of 4.07% across folds. Based on the combined confusion matrix from all five folds, the gendot class obtained a precision of 1.00 and a recall of 0.93, the raksasa class obtained a precision of 0.75, a recall of 0.69, and an F1-score of 0.72, and the rawit class obtained a precision of 0.71, a recall of 0.80, and an F1-score of 0.75, with most misclassifications occurring between the raksasa and rawit classes due to their morphological similarity. These findings show that the hybrid CNN–LVQ approach can be applied to chili seed classification without requiring specialized hardware or a pre-trained model.

DEVELOPMENT OF A WEB-BASED EMPLOYEE ASSET INVENTORY INFORMATION SYSTEM FOR KEMENTERIAN PELINDUNGAN PEKERJA MIGRAN INDONESIA

2026-09-22T07:42:56+00:00

Asset inventory management in government agencies requires a system that is able to ensure data accuracy, ease of monitoring, and presentation of information quickly and accurately. In the Kementerian Pelindungan Pekerja Migran Indonesia (KP2MI), the inventory process of assets used by employees is still carried out manually, potentially causing data duplication, recording errors, difficulty in tracking the whereabouts of assets, and delays in preparing reports. This study aims to develop a web-based employee asset inventory information system to improve the effectiveness, efficiency, and transparency of Barang Milik Negara (BMN) in the KP2MI environment. System development using the Waterfall method which includes the stages of needs analysis, design, implementation, testing, and maintenance. The system is developed using PHP programming language with MySQL database, while data collection is done through observation, interview, and literature study. Evaluation of the system is carried out using Black Box Testing to test the functionality of the system and the system Usability Scale (SUS) to measure the level of ease of use based on user perceptions. The test results showed that all 11 functional test scenarios were successfully executed, so that all system functions met the specified needs. In addition, the usability test obtained a SUS score of 82, which belongs to the category of Excellent with Grade A, indicating that the system has a very good level of ease of Use and can be accepted by users. The system developed is able to integrate asset data management, employee data, lending, returns, asset mutations, and presentation of reports in a centralized and real-time manner, thereby increasing data accuracy, accelerating report preparation, and supporting decision-making in asset management in the KP2MI environment.

DEEP LEARNING-BASED INTRUSION DETECTION SYSTEMS: ARCHITECTURES, IMPLEMENTATIONS, AND CHALLENGES IN MODERN NETWORK ANOMALY DETECTION

2026-09-23T07:43:27+00:00

The increasing complexity of cyber threats has strengthened the need for adaptive network intrusion detection systems (IDS). This systematic literature review (SLR) synthesizes nine peer-reviewed studies published from 2024 to 2026 on deep learning (DL)-based network anomaly detection. The review follows a PRISMA 2020-aligned process covering database selection, eligibility screening, quality assessment, and structured data extraction. Five research questions examine DL architectures, preprocessing and class-imbalance handling, comparative performance, deployment challenges, and the basis for a conceptual framework. The synthesis compares preprocessing, classification strategy, Accuracy, Precision, Recall, and F1-Score where reported. The evidence shows that hybrid architectures can achieve very high benchmark performance, but the results are dataset- and experimental-setting dependent; therefore, the claim that hybrid models universally exceed 95% is not supported. Across the included studies, the recurring technical pattern is: representative data collection, preprocessing and imbalance mitigation, architecture selection according to spatial or temporal characteristics, and multi-metric evaluation. The revised framework extends this four-layer pipeline by explicitly integrating Explainable AI (XAI) and edge-oriented deployment as cross-cutting operational requirements. The review also identifies limitations in dataset realism, cross-dataset validation, reporting consistency, computational efficiency, and explainability.

COMPARATIVE COMPLAINT TEXT CLASSIFICATION USING SVM, CNN, AND BERT WITH PRIORITY SCORING FOR NYC 311 SERVICE REQUESTS

2026-10-02T07:51:18+00:00

Abstract

Public complaint management systems like NYC 311 generate millions of service requests annually, posing significant challenges regarding manual categorization and resource prioritization. This study has two objectives: (1) to compare the classification performance of Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Bidirectional Encoder Representations from Transformers (BERT) on a large-scale city complaint dataset, and (2) to develop a complaint prioritization framework that combines normalized complaint frequency with average resolution time. The NYC 311 Service Request dataset, comprising 1,203,418 cleaned records, was utilized, with a sample of 20,000 records selected via stratified sampling. Pre-processing steps included converting text to lowercase, removing numbers and special characters, and normalizing whitespace. All models were evaluated without class weighting using a shared test set of 4,000 records. The results indicate that BERT achieved the highest macro F1-score (0.9167), outperforming SVM (0.8861) and CNN (0.8900), while all three models recorded similar aggregate accuracy rates (95.60%–95.83%). Both SVM and CNN yielded a recall of zero for the "Noise – Street/Sidewalk" category, whereas BERT was able to partially recognize it demonstrating that contextual language understanding is essential for distinguishing complaint categories with semantic ambiguity. Priority score analysis identified "Illegal Parking" (0.6007) and "HEAT/HOT WATER" (0.5437) as the highest-priority complaint types, while geographic analysis revealed that Brooklyn and the Bronx bear a disproportionately higher burden of complaints compared to other boroughs. These findings indicate that transformer-based classification combined with composite priority assessment can support data-driven municipal complaint management that accounts for geographical differences.

ARRHYTHMIA DETECTION FROM IMBALANCED ECG DATA USING CNN, FOCAL LOSS, AND RR INTERVAL FEATURES WITH INTER-PATIENT EVALUATION

2026-10-06T08:04:18+00:00

Detection of arrhythmias from electrocardiogram (ECG) signals is crucial for early diagnosis of heart rhythm disorders. However, the imbalanced distribution of arrhythmia classes poses challenges for classification models. This study proposes a Convolutional Neural Network (CNN) model combined with Focal Loss and RR features to address class imbalance using an inter-patient evaluation approach. The dataset includes three classes: N (Normal), SVEB (Supraventricular Ectopic Beats), and VEB (Ventricular Ectopic Beats), divided into training (DS1), validation (DS1-Val), and testing (DS2) sets. The proposed CNN + Focal Loss + RR model achieved an accuracy of 95.75% and an F1-score of 95.19%. Additionally, Explainable AI (XAI) using LIME is applied to improve model interpretability by identifying important features influencing predictions. The results demonstrate that the proposed approach provides accurate and interpretable arrhythmia detection, supporting reliable medical decision-making. However, the model's performance on the minority SVEB class remains limited, with a recall of only 34.24%, indicating that SVEB beats are frequently misclassified as Normal due to morphological similarities and severe class imbalance. This limitation highlights the need for further improvements in minority-class detection for clinical applicability.

APRIORI ALGORITHM AND TIME SERIES FOR PURCHASING PATTERNS AND SALES FORECASTING: MATAHARI CIPUTRA JAKARTA

2026-11-03T08:23:37+00:00

Retail transaction data can be used to explain purchasing patterns and forecast future sales, but the two analyses are often applied separately. This study examines purchasing patterns and product sales forecasts for the Eid shopping period at Matahari Department Store Ciputra Jakarta. The dataset contained 389,588 product-level records from January-April 2025 and January-April 2026, representing 194,854 transactions. The Apriori algorithm was applied with a minimum support of 1% and a minimum confidence of 10%. Daily sales of the 50 best-selling products were forecast using ARIMA(2,1,2), SARIMA(2,1,2)(1,1,1,7), and Prophet with weekly seasonality. Apriori produced 59 frequent itemsets and 24 association rules. The strongest selected rule was Socks to Sneakers with a lift of 11.39, while Hoodie to Jeans reached 43.63% confidence. ARIMA produced the lowest average errors, with MAE 42.46, RMSE 76.86, and MAPE 102.14%, compared with SARIMA and Prophet. The relatively high MAPE indicates that the forecast should be treated as a comparative estimate rather than an exact value. The 90-day forecast was combined with association rules to generate stock priorities, bundling, and product-display recommendations. Manual recalculation of selected Apriori rules and ARIMA evaluation metrics matched the system output, while all 16 black-box scenarios were successfully executed.

Keywords: Apriori; association rules; Eid sales; purchasing patterns; time series forecasting

XGBOOST-BASED CLASSIFICATION OF POTENTIALLY FRAUDULENT SHOPEE SELLERS USING SELLER METADATA AND SHAP

2026-11-03T08:23:06+00:00

The rapid growth of e-commerce increases the exposure to fraudulent sellers, while reactive reporting mechanisms generally operate only after consumers suffer losses. This study develops a proactive fraud-risk screening model to assist in identifying potential fraud-risk indications among Shopee sellers from store-profile metadata and to explain the model's decisions. Primary data were collected from Indonesian Shopee seller profiles using a manual-hybrid approach with Playwright. The initial dataset contained 460 profiles and was reduced to 457 clean records after duplicate removal. The final dataset comprised 355 Safe stores and 102 Fraud stores (labeled heuristically using external indicators) with 11 predictor features. XGBoost was trained with class weighting through a scale_pos_weight value of 3.48 and an 80:20 stratified split. Evaluation on 92 test records produced 95.65% Accuracy, 84.00% Precision, 100.00% Recall, and 91.30% F1-Score. SHAP analysis identified historical_sold (48.42%), store age (18.73%), and rating_star (14.76%) as the three most influential features. The trained model was subsequently integrated into a Chrome Extension prototype to provide predictions and SHAP-based rationales. The results indicate that a seller-centric metadata approach can support early fraud-risk detection; however, the 84.00% precision confirms that the output should be interpreted as a risk screening indication rather than an absolute determination of fraud, assisting consumers in making safer decisions.

IMPLEMENTATION OF DISASTER MITIGATION CHATBOT USING GRAPH RETRIEVAL-AUGMENTED GENERATION (GraphRAG) FOR BPBD KABUPATEN BOGOR

2026-11-03T08:23:20+00:00

Kabupaten Bogor is among the highest disaster-risk regencies in West Java, having recorded 1,826 disaster events throughout 2025, while official mitigation guidelines remain confined to static PDF documents that are difficult to consult during emergencies. To address this gap, this study implements a disaster mitigation chatbot built on Graph Retrieval-Augmented Generation (GraphRAG), pairing Neo4j Desktop as a local graph database with Gemma3:4b (served through Ollama) as the language model, orchestrated via LangChain and exposed through a Flask web interface. The knowledge graph was built from 11 source documents provided by BPBD Kabupaten Bogor—Disaster Preparedness Pocket Books together with supplementary disaster-information documents—spanning 8 disaster types, yielding 483 nodes and 754 relationships. Each entity node carries three properties (name, description, and disaster phase) to enable more precise retrieval. System performance was assessed with the RAGAS v0.4.3 framework over a 99-question dataset over three trials, yielding average Faithfulness, Answer Relevancy, and Context Precision scores of 0.8764, 0.6773, and 0.8216, respectively. These results indicate promising contextual grounding and retrieval performance, while the lower Answer Relevancy score highlights the need for further improvement in aligning generated answers with user questions. The findings demonstrate the technical feasibility of a locally hosted GraphRAG prototype for retrieving disaster-mitigation information from BPBD Kabupaten Bogor documents. However, the evaluation does not establish operational readiness for emergency use, as comparison with conventional RAG, expert assessment, and validation with intended users remain necessary.

CLEARPAY: INNOVATIVE DESIGN OF AN IMPORT INVOICE PAYMENT SYSTEM USING DESIGN THINKING

2026-11-03T08:22:52+00:00

The invoice payment process in import activities faces various obstacles, such as limited information, unorganized workflows, difficulty monitoring payment status, and the risk of administrative errors. The purpose of this research is to create a design for the invoice payment system in the import process by applying the Design Thinking method. This method consists of five stages, namely Empathise, Define, Ideate, Prototype, and Test. At the Empathise stage, information is collected through observation and interviews to understand user needs. Define is used to determine the main problems faced by users. Ideate is used to create various creative design solutions. The Prototype phase resulted in a high-fidelity prototype called ClearPay that provides features for invoice management, approval processing, payment processing, status monitoring, and payment confirmation, while Testing to assess the design based on user feedback. Evaluation using the system Usability Scale (SUS) obtained an average score of 82.42, which included Acceptable, Grade A, and the adjective rating Excellent categories. The results show that the ClearPay prototype has a good level of usability and user acceptance and can be the basis for system development at the implementation stage.