Journal of System and Computer Engineering

Crop Recommendation Based on Soil and Weather Conditions Using the K-Nearest Neighbors Algorithm

2025-08-02T14:51:16+00:00

The national food self-sufficiency program demands innovation in optimizing the selection of agricultural commodities based on environmental and weather conditions. This challenge is rooted in a fundamental problem faced by farmers—achieving harmony among soil characteristics, weather patterns, and suitable crops. In support of this initiative, it is necessary to develop a crop recommendation system based on machine learning that utilizes key soil and weather condition parameters. This study employs the K-Nearest Neighbors (KNN) algorithm, which functions by identifying the optimal value of ‘K’ to maximize classification accuracy. The KNN algorithm is implemented in a crop recommendation system to classify 1,100 datasets representing ideal growing conditions for 11 crop types. These datasets were generated using a normal distribution approach with a 5% variation from the mean values, and were validated using a clipping function to ensure the data remained within ideal ranges. The results of this study demonstrate that the KNN algorithm achieves high accuracy 96,67% in utilizing soil and weather parameters to generate crop recommendations. The average probability score for the recommended crops was 83.33%. Based on experimental testing, rice was recommended during the rainy and extreme rainy seasons, soybeans were recommended during the dry season, and mung beans were most suitable during extreme dry conditions.

Performance Exploration of Tree-Based Ensemble Classifiers for Liver Cirrhosis: Integrating Boosting, Bagging, and RUS Techniques

2025-08-03T15:11:50+00:00

Liver cirrhosis, as a significant chronic liver disease, exhibits a rising global prevalence, demanding more effective preventive approaches. In an effort to enhance early detection and patient management, this research proposes the development of a liver cirrhosis risk prediction model using machine learning technology, specifically comparing the performance of three ensemble tree models: Ensemble Boosted Tree, Ensemble Bagged Tree, and Ensemble RUSBoosted Tree. Utilizing clinical and laboratory data from adults with a history or risk of cirrhosis, the study reveals that Ensemble Bagged Tree achieved the highest accuracy at 71%, followed by Ensemble Boosted Tree (67.2%) and Ensemble RUSBoosted Tree (66%). Analysis of clinical and laboratory variables provides further insights into the most significant contributors to risk prediction. The findings lay the groundwork for the advancement of a more sophisticated liver cirrhosis risk prediction tool, supporting a vision of more personalized and effective preventive strategies in liver disease management

Graph-Based Fraud Detection with Optimized Features and Class Balance

2025-08-02T14:53:09+00:00

The increasing use of digital transactions also elevates the risk of fraud, particularly in credit card transactions. Fraud detection poses a challenge due to the highly imbalanced nature of the data and the complexity of relationships among entities. This study proposes a GNN-based approach, integrated with feature selection techniques and class imbalance handling through class weighting based on data distribution. Feature selection was performed using two methods: Correlation-based Feature Selection (CFS) and Random Forest Feature Importance, to obtain the most relevant features. Experimental results show that the combination of Random Forest feature selection and class weighting yielded the highest F1 Score, despite a slight decrease in accuracy. This indicates that feature selection and class weighting strategies can improve the model's ability to detect rare fraudulent transactions. This approach contributes to the development of more accurate and adaptive fraud detection systems in digital transaction environments.

Enhancing Intrusion Detection Using Random Forest and SMOTE on the NSL‑KDD Dataset

2025-08-02T15:06:38+00:00

Intrusion Detection Systems (IDS) play a crucial role in identifying suspicious activities on computer networks. However, a major challenge in developing machine learning-based IDS is the issue of class imbalance, where attacks—being minority classes—are often overlooked by classification models. This study aims to construct an intrusion detection system based on the Random Forest algorithm integrated with the Synthetic Minority Over-sampling Technique (SMOTE) to address this problem. The NSL-KDD dataset is used for evaluation, with the data split into 80% for training and 30% for testing. Experiments include Random Forest-based feature selection and performance evaluation using accuracy, precision, recall, and F1-score metrics. The results show that the Random Forest–SMOTE combination achieves an accuracy of 99.78%, precision of 99.70%, recall of 99.88%, and an F1-score of 99.79%. The confusion matrix indicates a very low rate of false positives and false negatives. Additionally, selecting the most influential features such as src_bytes and dst_bytes improves model efficiency. Thus, the integration of Random Forest and SMOTE proves to be effective in enhancing detection sensitivity toward attacks without compromising model precision. This approach offers a significant contribution to the development of adaptive, accurate, and deployable IDS in real-world network environments.

Decision Support System for Selecting Used Cars Using the Analytical Hierarchy Process (AHP) Method Based on a Website at CV Auto Mobil Manokwari

2025-08-02T15:09:25+00:00

Buying a used car is often considered by the public as an alternative because it is more affordable than a new one. However, the process of choosing a used car is not easy because there are various factors that must be considered, such as engine condition, completeness of documents, physical condition, price, engine capacity, and year of manufacture. At CV Auto Mobil Manokwari, prospective buyers often have difficulty determining the choice of a used car that best suits their needs and budget. This research aims to design a website-based decision support system using the Analytical Hierarchy Process (AHP) method to assist buyers in choosing used cars objectively and systematically. The AHP method is used to compare each criterion in pairs and determine the priority weight of each criterion. The system was developed using the PHP programming language and MySQL database with a waterfall approach. With this system, the process of selecting used cars becomes more directed, accurate, and efficient, as well as helping users make decisions practically and quickly, and objectively.

Comparative Analysis of SVM and IndoBERT for Intent Classification in Indonesian Overtime Chatbots

2025-08-04T13:26:35+00:00

Digital transformation in higher education requires the development of intelligent and adaptive information systems, including services such as overtime submission for university staff. Chatbots offer a promising solution to enhance user interaction with the E-LEMBUR system. However, developing chatbots in academic settings poses challenges, including limited training data, complex overtime policies, and diverse institutional terminology. This study compares two intent classification approaches: Support Vector Machine (SVM), a traditional machine learning method, and IndoBERT, a transformer-based model designed for the Indonesian language. The dataset comprises 250 real user queries from the overtime system at Institut Teknologi Sepuluh Nopember (ITS). Experimental results show IndoBERT achieves 87% accuracy, slightly outperforming SVM at 85%. While IndoBERT offers better accuracy, it demands higher computational resources, presenting a trade-off between performance and efficiency. This study contributes by validating IndoBERT’s effectiveness on a limited dataset, establishing an initial benchmark for intent classification in overtime chatbots, and offering implementation recommendations aligned with university IT infrastructure. These findings lay the groundwork for developing context-aware information systems for staff services in Indonesian higher education.

Implementation of the K-Means Algorithm for Clustering Students’ Web Programming Course Grades Using Silhouette Score

2025-08-05T23:40:24+00:00

The development of information technology requires students majoring in informatics engineering to master web programming as one of the core competencies of the study program. Variations in students' ability to understand the material are reflected in significant differences in grades, so an objective analysis approach is needed to determine the ability of students. This study aims to group students based on academic grades in Web Programming courses using the K-Means algorithm. The data analyzed includes 1-3 assignment grades, attendance, UTS, and UAS from 32 students in the Department of Informatics Engineering, University of Papua. The research stages include preprocessing, data normalization, and clustering process using Orange Data Mining tools. Determination of the optimal number of clusters is done using the Silhouette Score method, and the best results are obtained at K = 4 with a Silhouette Score value of 0.513 which indicates a good cluster structure. The clustering results show that Cluster 1 has the highest score with a final score ranging from 0.93-1 with an Excellent score category consisting of 8 students, Cluster 2 with a Poor score category consists of 10 students with a final score range of 0.23-0.61, then Cluster 3 with a Good score category consists of 10 students with a Final score of 0.78-0.87 and Cluster 4 with a Fair score category consists of 4 students with a score range of 0.64-0.75. The results of this study provide information about the distribution of student abilities and can be the basis for improving learning strategies in the future.

Enhancing Flood Prediction Using Hybrid LSTM-Transformer Deep Learning Approach

2025-08-07T05:21:45+00:00

Flood prediction is crucial for effective disaster management, yet it remains a complex challenge due to the nonlinear nature of meteorological processes. This study develops and evaluates a novel hybrid model that integrates Long Short-Term Memory (LSTM) networks and Transformer attention mechanisms to enhance predictive accuracy for rainfall-based flood forecasting. Using extensive Australian weather data collected from 49 stations over a decade (2007-2017), the model incorporates comprehensive feature engineering, including derived meteorological indicators, rolling statistical measures, and temporal lag features. The hybrid LSTM-Transformer architecture achieved superior precision (77.69%) and high accuracy (84.57%) compared to a Random Forest baseline model. Confusion matrix analysis illustrated the hybrid model’s strength in reducing false alarms, indicating a conservative yet highly reliable predictive performance. Feature correlation analysis revealed important relationships among temperature, humidity, pressure, and rainfall, highlighting the complexity of meteorological interactions. The findings demonstrate the effectiveness of integrating sequential and global temporal modeling for flood prediction, providing valuable guidance for operational forecasting systems and disaster preparedness strategies. This research contributes significantly to existing flood forecasting methodologies and suggests promising directions for future enhancements.