Jurnal KomtekInfo

Implementation of a Hybrid MobileNetV2 Model for Early Detection of Breast Cancer Based on Mammography Images from RSUP Dr. M. Djamil Padang

Lily Aulia — 2026-09-30

Breast cancer poses a significant health risk to women; however, early detection through mammographic imaging continues to face challenges in achieving high diagnostic accuracy using conventional diagnostic methods. Deep learning offers a solution for improving the accuracy of medical image classification; however, a single model remains limited, so a hybrid approach combining two methods is needed to achieve optimal performance. This study aims to develop and evaluate a Hybrid Deep Learning model that combines MobileNetV2 and XGBoost to detect and classify breast cancer based on mammography images. The research method begins with image preprocessing consisting of resizing, normalization, and augmentation, after which the dataset is split into training and testing data with an 80:20 ratio. MobileNetV2 is used as a feature extractor, producing a 1,280-dimensional feature vector, which is then classified by XGBoost into three classes: normal, benign, and malignant. The dataset consists of 224x224-pixel breast cancer mammography images sourced from RSUP Dr. M. Djamil Padang and the Kaggle platform. The test results show that the Hybrid MobileNetV2 and XGBoost model achieved an accuracy of 75%, a weighted precision of 77%, a weighted recall of 75%, and a weighted F1-score of 74%. The model performed best on the normal class, while the benign class experienced misclassification due to visual similarity with other classes. This study is expected to serve as a foundation for developing an accurate and efficient digital image-based breast cancer detection system to support medical personnel in clinical decision-making.

Data Mining Analysis Using K-Means and Naive Bayes Methods for Drug Inventory Prediction

Filsha Rifi Maharani — 2026-09-30

The management of drug inventory still encounters issues related to unstable stock fluctuations, such as drug shortages and overstocking. This condition has the potential to cause operational losses and degrade the quality of patient care. This study aims to analyze categorized drug inventory using the K-Means Clustering and Naive Bayes algorithms. The K-Means algorithm is employed to group drug data into three categories: low, sufficient, and excessive. The Naive Bayes algorithm is utilized to classify and predict the status of drug inventory requirements. The dataset was obtained from RSUD Tualang over the past two years, comprising 206 drug data points with variables including name, category, stock, price, demand, and expiration date. The test results indicate that the K-Means algorithm successfully grouped the data into three clusters, namely Low (14 items), Sufficient (125 items), and Excessive (67 items), while the Naive Bayes algorithm successfully classified the status of drug inventory requirements with an accuracy of 95.63%. The findings demonstrate that the combination of Data Mining utilizing K-Means and Naive Bayes is effective for drug inventory analysis at RSUD Tualang, thereby assisting management in identifying inventory status and supporting more accurate, efficient, and timely procurement decision-making compared to conventional methods previously employed.

Clustering of Bandwidth Usage Patterns of Regional Government Organizations in Solok City Using a Combination of Self-Organizing Map and K-Means

Hikko Razaki — 2026-09-30

Differences in the intensity of digital services across Regional Government Organizations (OPDs) within the Solok City Government result in varying bandwidth requirements among institutions. Uniform capacity allocation does not adequately reflect these differences and may lead to network inefficiencies. This study classified bandwidth-usage patterns using a combination of the Self-Organizing Map (SOM) and K-Means Clustering to provide a basis for more proportional bandwidth allocation. The dataset comprised 1,326 daily records from 26 OPDs collected between March and May 2026, including upload volume, download volume, number of employees, and number of active applications. Data processing involved outlier detection, aggregation, Min-Max normalization, SOM training on a 5 × 5 grid, determination of the optimal number of clusters using the Elbow Method, selection of initial centroids from the extreme points of the SOM map, and final clustering using K-Means. The SOM mapping produced a Quantization Error of 0.040207 and a Topological Error of 0.076923. Three clusters were identified as the optimal configuration. The SOM–K-Means model achieved a Silhouette Score of 0.448335, a Davies–Bouldin Index of 0.817415, and a WCSS value of 2.702929, outperforming conventional K-Means. The final results classified 11 OPDs into the high-usage category, 4 into the moderate-usage category, and 11 into the low-usage category. These findings provide an empirical basis for allocating bandwidth according to the usage characteristics of each OPD

Integrating Educational Data Mining through Decision Tree and Random Forest Comparison for Assessing Employment Suitability

Riski Hidayatullah Maas — 2026-09-30

Educational Data Mining can be utilized to identify patterns in educational data and support the evaluation of the alignment between graduates’ competencies and labor market requirements. Assessing alumni employment suitability is essential for providing an overview of their readiness and job alignment after completing their education. This study aims to compare the performance of the Decision Tree (DT) and Random Forest (RF) algorithms in classifying alumni employment suitability levels. The research stages included data collection, selection of employed alumni data, data cleaning, transformation of categorical variables into numerical scores, target label construction, division of the dataset into training and testing sets, and class balancing using Random Over-Sampling. The classification models were constructed using the DT and RF algorithms and evaluated based on accuracy, precision, recall, F1-score, and confusion matrix. The dataset was obtained from the 2025 Alumni Tracer Study conducted by Universitas Negeri Padang and comprised 2,831 records with eleven variables related to alumni employment and competencies. The results indicate that the Decision Tree achieved an accuracy of 83.42%, a macro-averaged precision of 76.74%, a macro-averaged recall of 76.82%, and a macro-averaged F1-score of 76.59%. Random Forest demonstrated better performance, achieving an accuracy of 90.83%, a macro-averaged precision of 83.85%, a macro-averaged recall of 82.22%, and a macro-averaged F1-score of 82.94%. This study contributes an Educational Data Mining-based approach that can assist higher education institutions in evaluating alumni employment suitability and provide a basis for formulating policies aimed at improving graduates’ competencies.

Deep Learning-Based Image Classification of B.1-KWK-Perseorangan Forms Using K-Means and Convolutional Neural Networks

Sandi Mahardika — 2026-09-30

The verification process of B.1-KWK-Perseorangan forms at General Election Commission (KPU) is still conducted manually, resulting in time-consuming procedures and increasing the risk of document misclassification. The application of deep learning provides a promising solution for improving the efficiency and accuracy of administrative election document image classification. This study aims to analyse the capability of Convolutional Neural Networks (CNN) in classifying images of B.1-KWK-Perseorangan forms and to investigate the effect of K-Means as an initial image segmentation stage on improving classification accuracy and the effectiveness of document verification. The proposed method employs a deep learning approach by integrating K-Means for image segmentation and CNN for image classification. The research workflow consists of image preprocessing, segmentation, model training, and performance evaluation. The dataset comprises digital images of B.1-KWK-Perseorangan forms from the 2024 election, provided by the KPU of Sungai Penuh City, with an image resolution of 512 × 783 pixels. The experimental results demonstrate that the proposed model achieved an accuracy of 94% and a macro-average F1-score of 0.95, indicating that the integration of K-Means and CNN is effective for the classification of B.1-KWK-Perseorangan form images. Furthermore, the developed model has the potential to improve the speed, consistency, and accuracy of document verification within the KPU, while supporting the development of more effective, accurate, and efficient Deep Learning-based classification systems for election administrative documents.

Predictive Analysis of Students’ Academic Readiness for the Academic Competency Test Using an Integrated K-Means and C4.5 Approach

Sri Atiqah Elvidamayanti — 2026-09-30

The identification of students' academic readiness for the Academic Ability Test (Tes Kemampuan Akademik/TKA) is still carried out conventionally and subjectively, causing academic interventions to be frequently delayed and poorly targeted. In fact, the academic data already available in schools holds considerable potential for early, objective, data-driven detection of student readiness patterns. This study aims to build a predictive-analysis model of students' academic readiness for the TKA by integrating the K-Means and C4.5 methods. K-Means is applied to group students according to the similarity of their academic characteristics, after which the resulting clusters are used as class labels to build a decision-tree classification model using C4.5. The dataset was obtained from State Senior High School (SMAN 1) Kamang Magek, comprising 82 students with 11 predictor variables covering academic aspects (average compulsory and elective subject scores, score trend, score stability, attendance) and non-academic aspects (learning motivation, parental support, learning intensity, mental condition). The resulting model achieved an accuracy of 93.90% with a weighted F1-score of 0.93, with K-Means identifying three readiness groups (Ready, Fairly Ready, Not Ready) consisting of 45, 20, and 17 students, respectively. Learning motivation proved to be the most dominant factor in the C4.5 decision tree, followed by score trend, score stability, and attendance. This study contributes an interpretable early-detection model of academic readiness that can be operationalized as a basis for designing data-driven academic intervention programs in senior high schools.