Explorer

Optimasi Kinerja Model LeNet Berbasis Deep Learning untuk Klasifikasi Citra Menggunakan Pendekatan Hyperparameter Tuning

2026-07-20T03:01:40+00:00

Abstract−The development of Deep Learning has made significant contributions to the field of image classification, particularly through the use of Convolutional Neural Networks (CNN). However, one of the main problems in implementing CNN models is the suboptimal performance of the model due to inappropriate hyperparameter selection. Simple models such as LeNet are often considered to have limited performance compared to modern architectures, even though with the right approach, this model still has the potential to produce competitive performance. Therefore, this study aims to improve the performance of the LeNet model in image classification through a hyperparameter tuning approach. The methods used in this study include data preprocessing, dataset division, implementation of the LeNet model as a baseline, and hyperparameter optimization including learning rate, batch size, optimizer, and number of epochs. The dataset used is a handwritten number image that has been normalized and transformed to match the CNN model input. Model performance evaluation was carried out using accuracy, precision, recall, and F1-score metrics. The results showed that hyperparameter optimization provided a significant performance improvement over the LeNet model. The baseline model produced an accuracy of 97.52%, while the best optimized model achieved an accuracy of 98.93%. Furthermore, precision, recall, and F1-score values also improved, indicating the model's improved and more balanced classification capabilities. Thus, this study demonstrates that hyperparameter optimization is an effective approach to improving the performance of simple CNN models without increasing architectural complexity.

Integrasi Metode AHP-SAW untuk Evaluasi Kinerja Mitra Bisnis pada Sistem Manajemen Kemitraan Berbasis Web

2026-07-23T02:43:41+00:00

Business partnership management in SMEs and companies is still largely conducted manually, making objective partner performance evaluation difficult. This study aims to integrate the Analytical Hierarchy Process (AHP) and Simple Additive Weighting (SAW) methods to evaluate partner performance in a web-based partnership management system. AHP is used as the criteria weighting method through pairwise comparison validated by the Consistency Ratio (CR), while SAW is used for partner ranking. Four evaluation criteria were established: number of collaboration activities, active partnership duration, completeness of collaboration evidence, and number of expired MoUs without renewal. The AHP calculation produced criteria weights with a CR value of 0.0038 (< 0.1), indicating consistent judgment. Testing on five partners produced consistent rankings, with Partner C as the best-performing partner (preference value 1.0000). A comparative analysis between AHP-SAW and pure SAW (manual weighting) showed identical rankings (Spearman correlation = 1.000), yet AHP-SAW provided more objective and validated weighting justification. The developed system proved capable of producing objective, transparent, and measurable partner performance evaluations.

The Multimodal Medical Clustering of Chest X-Ray Images and Clinical Reports Using Deep Visual and Semantic Text Features

2026-07-24T04:59:55+00:00

The limited availability of labelled medical data is a real obstacle in building a supervised learning-based Radiology analysis system. This study offers an alternative approach, unlabeled multimodal clustering, that simultaneously combines clinical information sources: chest X-ray images and physician clinical reports. Visual features are extracted from 1,000 images using a pretrained ResNet152, and text features are constructed using a Dual TF-IDF Vectorizer that processed the findings and conclusions columns separately with explicit weightings (0.4 and 0.6), then enriched with 10 related topics from the Latent Dirichlet Allocation (LDA) model. These two modalities are combined through a weighted late fusion strategy after being normalized by L2, with a visual weight of 0.5 and a text weight of 0.5, respectively. Before clustering, dimensionality reduction is performed using UMAP with 50 components and a cosine metric. Determining the optimal K value using the Elbow and Silhouette Score methods in the range of K=2 to K=15 showed K=8 as the best choice with a Silhouette Score value of 0.7524. The clustering results reveal eight clinical groups reflecting diverse diagnostic patterns, ranging from normal to abnormal findings, such as cardiomegaly and pneumonia, that require clinical attention. This approach has the potential to serve as the basis for a medical data exploration system that relies on no label annotations.

Monitoring Suhu dan Kelembapan Ruang Server Berbasis IoT Menggunakan ESP32, DHT22, Blynk, dan ThingSpeak

2026-07-25T16:24:09+00:00

Penelitian ini bertujuan merancang dan mengimplementasikan sistem pemantauan suhu dan kelembaban ruang server berbasis Internet of Things (IoT) menggunakan mikrokontroler ESP32 dan sensor DHT22. Sistem ini terintegrasi dengan Blynk untuk pemantauan real-time, ThingSpeak untuk penyimpanan data historis, serta notifikasi otomatis melalui Telegram dan indikator LED lokal berdasarkan ambang batas 30°C. Penelitian mengikuti model pengembangan prototipe, meliputi perancangan perangkat keras dan perangkat lunak, simulasi, integrasi, dan pengujian langsung di ruang server STKIP PGRI Situbondo. Hasil pengujian menunjukkan sistem mampu membaca data lingkungan secara akurat, mentransmisikannya ke platform cloud, dan menampilkan informasi secara lokal melalui layar OLED. Fluktuasi suhu selama operasional server tercatat dengan baik, dengan puncak awal mencapai 34,2°C dan stabil pada rentang 20–26°C. Sistem berhasil mengaktifkan notifikasi dan indikator LED hanya ketika ambang batas terlampaui, sehingga tidak terjadi peringatan palsu pada kondisi normal. Kesimpulannya, sistem pemantauan berbasis IoT ini menyediakan pemantauan real-time, pencatatan data historis, dan mekanisme peringatan dini yang efektif. Sistem ini mendukung pemeliharaan ruang server secara proaktif. Pengembangan selanjutnya dapat mencakup penambahan sensor redundan, saluran komunikasi cadangan, atau mekanisme kontrol pendinginan otomatis untuk meningkatkan keandalan dan fungsionalitas sistem.

Integrasi Strategi Pre-processing Data untuk Optimalisasi Akurasi Algoritma Backpropagation

2026-07-25T17:20:49+00:00

Backpropagation is one of the artificial neural network algorithms widely used in classification and prediction processes due to its ability to recognize data patterns accurately. However, the performance of this algorithm is highly influenced by the quality of the input data. Unstructured data, differences in data scales, missing values, and irrelevant features can reduce the model’s accuracy. This study aims to analyze the effect of integrating data pre-processing strategies to optimize the accuracy of the Backpropagation algorithm. The dataset used in this research was obtained from the Badan Pusat Statistik (BPS) in the form of Open Unemployment Rate data for the population aged 15 years and above in North Sumatra Province from 2019 to 2024. The applied pre-processing stages included data cleaning, normalization, missing value handling, and feature reduction. The research method was conducted by comparing the model testing results using standard pre-processing and partial pre-processing on several network architectures. The results showed that the implementation of pre-processing strategies was able to improve the performance of the Backpropagation model. The highest accuracy value was obtained in the 3-56-1 architecture with an increase from 80.00% to 85.88%. In addition to improving accuracy, the model training process became more stable and the error convergence was achieved faster. Therefore, the integration of data pre-processing strategies has proven to be effective in optimizing the accuracy of the Backpropagation algorithm for numerical data-based prediction problems

Klasifikasi Sentimen Teks Code-Mixed Indonesia–Inggris Non-Formal Pada X Menggunakan Model Fine-Tuned DistilBERT

2026-07-25T17:56:15+00:00

The rapid growth of social media has increased the use of Indonesian–English code-mixed language in digital communication, particularly on social media X. The non-formal characteristics of social media text, such as slang, abbreviations, emojis, and language switching within a single sentence, make sentiment analysis more challenging than monolingual text. This study aims to perform sentiment classification on code-mixed text by evaluating the performance of the lightweight Transformer model DistilBERT and comparing it with IndoBERTweet. The study adopts the CRISP-DM methodology, which consists of Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and Deployment stages. The dataset comprises 1,108 primary data collected from social media X between 2022 and 2026 and 5,048 secondary data obtained from previous research. Three experimental scenarios were applied: translation into Indonesian, translation into English, and raw data without translation. The results show that DistilBERT achieved its best performance on the English translation scenario, with accuracies of 73.66% on the primary dataset and 72.00% on the secondary dataset. Meanwhile, IndoBERTweet obtained the highest performance on the Indonesian translation scenario, achieving an accuracy of 79.01%. These findings indicate that the alignment between the language of the input data and the pre-training characteristics of the model significantly affects sentiment classification performance on non-formal code-mixed text.

Optimasi Support Vector Machine Menggunakan Particle Swarm Optimization pada Analisis Sentimen Ulasan Shopee COD

2026-07-26T16:43:41+00:00

The Cash on Delivery (COD) service provided by the Shopee e-commerce platform often elicits a large volume of user reviews that exhibit unconventional language structures, prompting the need for a precise and automated sentiment analysis mechanism. This research endeavor seeks to categorize sentiments expressed in Shopee reviews as either positive or negative by leveraging the Support Vector Machine (SVM) algorithm, which has been fine-tuned using Particle Swarm Optimization (PSO). A key obstacle in text analysis lies in the vast feature space, which can impair model efficacy. Thus, PSO is utilized as a feature selection technique to identify the most pertinent set of terms from the TF-IDF feature extraction. The findings reveal that the integration of PSO successfully decreased feature dimensionality by 45% from the initial set of 1,000 features. Despite the substantial reduction in features, the SVM-PSO model achieved an enhanced accuracy of 81.21%, surpassing the baseline model's 78.79%. With an AUC value of 0.845, it is evident that the model retains stability and effectiveness in discerning sentiment even with a considerably reduced feature set. This investigation illustrates the efficacy of PSO optimization in eliminating extraneous features and refining the model's focus on sentiment-carrying vocabulary.

Prototipe Alat Pendeteksi Banjir Berbasis Mikrokontroler dengan Fuzzy Tsukamoto

2026-07-27T12:54:37+00:00

Floods are among the most frequent natural disasters in Indonesia, causing significant material losses and casualties. Therefore, an early warning system capable of detecting flood potential quickly and accurately is required. This study aims to design and implement a flood detection prototype based on the ESP32 microcontroller using the Tsukamoto fuzzy logic method. The system employs an HC-SR04 ultrasonic sensor to measure water levels and an FC-37 rain sensor to detect rainfall intensity. Data collected from both sensors are processed using the Tsukamoto fuzzy logic approach through fuzzification, inference, and defuzzification stages to determine flood risk status categorized as safe, alert, and danger. The decision results are displayed through LED indicators and a buzzer, while real-time notifications are sent via the Telegram application using the ESP32 Wi-Fi connectivity. The testing results demonstrate that the system successfully classifies flood conditions according to the predefined fuzzy rules and automatically delivers notifications to users. The implementation of the Tsukamoto fuzzy logic method enhances decision-making capabilities in handling uncertain environmental conditions. Therefore, the developed prototype can serve as an effective, simple, and practical flood early warning system for flood-prone areas.

Explainable Diabetes Classification Using Machine Learning Algorithm and SHAP Analysis on Clinical Laboratory Data

2026-07-28T02:18:29+00:00

Diabetes mellitus is a chronic metabolic disorder that continues to pose significant health challenges worldwide due to its increasing prevalence and potential complications. Early and accurate identification of diabetes is essential to support timely intervention and effective disease management. Recent advances in machine learning have enabled the development of intelligent classification systems; however, many predictive models still suffer from limited interpretability, reducing their applicability in clinical environments. Therefore, this study proposes an explainable diabetes classification framework using machine learning algorithms and SHapley Additive exPlanations (SHAP) analysis on clinical laboratory data. The dataset consists of 1,000 patient records containing demographic and laboratory attributes, including age, gender, glycated hemoglobin (HbA1c), cholesterol, triglycerides, lipoprotein levels, creatinine, urea, and body mass index (BMI). Four machine learning algorithms, namely Decision Tree, Random Forest, Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost), were developed and evaluated. The performance of each model was assessed using Accuracy, Precision, Recall, F1-Score, Classification Report, and Confusion Matrix. Experimental results indicate that Random Forest and XGBoost achieved the highest classification accuracy of 98.5%, outperforming Decision Tree (98.0%) and SVM (94.5%). Due to its strong predictive capability and compatibility with explainable artificial intelligence techniques, XGBoost was selected for further SHAP analysis. The SHAP results successfully identified the most influential features contributing to diabetes classification and provided transparent explanations regarding model predictions. The proposed framework demonstrates that combining machine learning algorithms with explainable artificial intelligence improves predictive accuracy and interpretability, supporting reliable clinical decision-making for diabetes diagnosis.

Penerapan Algoritma K-Means pada Pengelompokkan Dampak Bermain Game Online Terhadap Minat Belajar

2026-07-30T08:30:09+00:00

This study aims to group students' level of learning interest based on the intensity of playing online games using the K-Means Clustering algorithm. The issue being addressed is the increasing activity of gaming, which can potentially affect students' learning behavior, but there hasn't been a structured mapping of student characteristics yet. The data used comes from questionnaires filled out by 65 respondents with 6 main variables, including playing frequency, playing duration, playing time, and learning interest indicators. The method used is K-Means with steps of preprocessing, normalization using StandardScaler, and testing the number of clusters using the Elbow method. The study results show that the optimal number of clusters is 3, with a silhouette score of 0.323 and a Davies-Bouldin Index of 1.465. It produced three groups, namely: (1) low learning interest, (2) medium learning interest, and (3) high learning interest. The clustering results showed that the majority of students were in the Medium Learning Interest category with 37 students (56.92%), followed by Low Learning Interest with 16 students (24.62%), and High Learning Interest with 12 students (18.46%). The silhouette score yielded a value of 0.323303, indicating a fairly good cluster structure. This study shows that most students are in a medium condition, meaning they still play games without significantly affecting their learning interest. The contribution of this research is providing a data-based approach to categorize students' learning interest levels related to online gaming activities, and it also serves as a basis for schools to design more effective monitoring and educational strategies. The research provides a mapping of student characteristics based on data that can be used as a basis for making decisions in academic guidance.