JURNAL FASILKOM

Analisis Pola Lalu Lintas Kapal Selat Bali Berbasis AIS, K-Means, Traffic Flow

Deny Adi Setyawan, Agustina Purwatiningsih, Febrian Sulistyo Budi — Sat, 10 Jan 2026 00:00:00 +0700

The Bali Strait is one of the busiest sea crossing routes in Indonesia, characterized by high intensity of ship movements and dynamic traffic patterns throughout the day. These conditions require comprehensive analysis to understand the characteristics of vessel movements, identify density zones, and determine peak periods that potentially increase navigation risks. This study aims to analyze ship traffic patterns in the Bali Strait using a combination of K-Means Clustering and Traffic Flow Model based on Automatic Identification System (AIS) data. The dataset consists of 790 AIS records collected during the period of 20–26 June 2025. The research stages include data preprocessing, determination of the optimal number of clusters using the Elbow method, classification of vessel movement behavior using the K-Means algorithm, and analysis of traffic parameters comprising volume, speed, density, and traffic flow. The results reveal the formation of three main clusters: low-speed vessels concentrated around port areas, medium-speed vessels operating on main trajectories, and high-speed vessels dominating the crossing lanes. Evaluation of clustering quality using the Silhouette Coefficient produced a value of 0.3040, indicating a reasonably good level of cluster separation. Furthermore, a consistent peak hour pattern was identified at 12:00, along with two high-density zones located near Ketapang Port and Gilimanuk Port. These findings demonstrate that AIS-based analysis is capable of providing measurable representation of the dynamics of ship traffic in the Bali Strait and has the potential to support operational optimization, enhancement of navigation safety, and consideration for the implementation of a Traffic Separation Scheme (TSS)

Klasifikasi Algoritma Kriptografi pada Pesan Terenkripsi menggunakan Support Vector Machine (SVM)

Sun, 11 Jan 2026 00:00:00 +0700

Data protection has become a highly critical aspect, particularly in addressing ransomware threats that illegally encrypt data. This study is important to evaluate the capability of machine learning techniques in identifying encryption algorithms used in encrypted data, especially in ransomware attacks. This work represents an initial step that can assist cybersecurity practitioners in more rapidly understanding attack patterns, determining appropriate response strategies, and enhancing proactive mitigation and response efforts to protect data against increasingly complex cyber threats. The machine learning algorithm employed in this study is the Support Vector Machine (SVM). The dataset consists of ciphertext generated using the AES, DES, and Vigenère Cipher cryptographic algorithms. The feature extraction process utilizes ten statistical features to capture the distinctive patterns of each type of ciphertext. The SVM model is developed using a data split of 90% for training and 10% for testing. Performance evaluation is conducted using a confusion matrix with accuracy, precision, recall, and F1-score metrics. The result demonstrate an average accuracy 0f 92,33%, with the vigenere cipher being perfectly classified (100% accuracy). Howefer, slight misclassifications occured beetween AES and DES duet o their similiar entropy chraracteristic. Experimental results demonstrate that the SVM model is capable of identifying encryption algorithms with high accuracy and balanced classification performance across the three algorithm classes. These findings highlight the potential of machine learning approaches for detecting encryption algorithms in cyber-attacks, thereby making a meaningful contribution to the improvement of proactive data security mitigation and response strategies.

Integrasi Metode Forward Chaining dan Teorema Bayes Untuk Identifikasi Diagnosa Penyakit Kulit Pada Kucing

Ade Anugerah, Sucipto, Syarifah Putri Agustini Alkadri — Wed, 31 Dec 2025 00:00:00 +0700

Skin diseases in cats are among the most common health issues, yet many cat owners still lack awareness of their symptoms. Limited access to veterinary services, especially in regions such as West Kalimantan, poses a significant challenge in early identification and treatment. This study aims to develop a web-based expert system capable of automatically diagnosing skin diseases in cats based on symptoms inputted by users. The system utilizes the Forward Chaining method for rule-based inference and the Bayes Theorem for probabilistic calculation to determine the likelihood of diseases. The system was built using the Laravel framework and MySQL database, based on a total of 83 case data obtained through direct interviews with veterinary experts. Testing using black-box and user acceptance methods showed that the system functions effectively and delivers accurate and informative diagnostic results. The system achieved an accuracy rate of 100% when tested on validated expert data. Therefore, this system can serve as an effective tool for cat owners to quickly and independently gain initial insights into their cat’s skin health before consulting a veterinarian.

Inovasi Agen AI Dalam Sistem Pencatatan Struk Digital Otomatis Berbasis n8n

Wahyu Pratama, Fadhli Almu'iini Ahda — Tue, 06 Jan 2026 00:00:00 +0700

The utilization of workflow automation and multimodal artificial intelligence introduces a new approach to developing an intelligent digital receipt recording system. This study aims to design an automatic transaction processing system by integrating n8n as a workflow engine, Google Gemini AI as a multimodal inference model, and Telegram Bot as a conversational interface. The system is implemented in a self-hosted Docker-based environment to ensure local execution without cloud dependence, enhancing data security and reducing operational costs. An experimental software engineering method was applied using 33 test scenarios consisting of 20 image inputs and 13 text inputs. The system successfully extracted key transaction information such as store name, total amount, and transaction date under various real-world conditions, including blurred images, faded ink, missing text segments, tilted receipts, and imperfect handwriting. Evaluation using a Confusion Matrix produced perfect classification results with 100% accuracy, precision, recall, and F1-score, confirming that all system outputs aligned with actual conditions. The system also demonstrated stable performance with average processing times of 15.8 seconds for text and 17–18.5 seconds for low-resolution images. These results indicate that combining workflow automation and multimodal AI provides an effective and adaptive solution for automatic transaction recording.

Peramalan Harga Emas (XAU/USD) menggunakan metode Sigle Exponential Smoothing (SES) dan Autoregressive Integrated Moving Average (ARIMA)

Balqis Aidil Adha, Mariza Devega — Fri, 09 Jan 2026 00:00:00 +0700

Gold (XAU/USD) is one of the most significant global commodities, often viewed as a safe-haven asset amid economic and political uncertainty. Accurate forecasting of gold prices is crucial for investors and policymakers in formulating strategic financial decisions. This study aims to compare the performance of the Single Exponential Smoothing (SES) and Autoregressive Integrated Moving Average (ARIMA) methods in forecasting gold prices using historical datasets from Kaggle, Investing.com, and ForexSB covering the period from January 2020 to September 2024. The analysis was conducted using Python on Google Colaboratory with evaluation metrics including Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The results show that both SES and ARIMA effectively captured the upward trend of gold prices, with SES achieving slightly better accuracy across all datasets. The lowest MAPE value of 0.62% was obtained using SES on the ForexSB dataset, indicating an excellent forecasting performance. Therefore, SES is considered more efficient and reliable for non-seasonal time series with stable trends

Analisis Sentimen Kebijakan Makan Bergizi Gratis Menggunakan IndoBERT dan Machine Learning

Danang Arbian Sulistyo, Erik Setiadi — Wed, 31 Dec 2025 00:00:00 +0700

Media sosial telah menjadi forum vital untuk opini publik terhadap kebijakan pemerintah, seperti program "Makan Bergizi Gratis" (MBG) di Indonesia. Memahami sentimen ini sangat penting bagi pemangku kepentingan. Penelitian ini bertujuan untuk (1) menganalisis distribusi sentimen publik terhadap kebijakan MBG dan (2) menentukan model machine learning terbaik untuk klasifikasi sentimen tersebut. Penelitian ini menggunakan 12.389 tweet yang dikumpulkan dari platform X. Metode hybrid labeling, yang mengkombinasikan leksikon berbasis domain dengan IndoBERT, diterapkan untuk melabeli data secara otomatis. Untuk klasifikasi, tiga model (Random Forest, XGBoost, dan Ensemble) dibandingkan menggunakan fitur hybrid (TF-IDF trigram, embedding IndoBERT, dan fitur leksikon) pada dataset yang telah diseimbangkan dengan SMOTE. Hasil penelitian menunjukkan bahwa sentimen publik didominasi oleh sentimen negatif (68,6%), diikuti oleh positif (19,5%) dan netral (11,9%). Model Random Forest menunjukkan kinerja tertinggi, dengan pencapaian F1-Score rata-rata 0.9383 pada K-Fold cross-validation dan 0.9363 pada test set final. Studi ini membuktikan bahwa pendekatan hybrid yang diusulkan sangat efektif untuk klasifikasi sentimen publik berbahasa Indonesia pada domain kebijakan pemerintah.

Perbandingan Metode Q-Learning Dan SARSA Dalam Optimasi Prediksi Tren Saham Pada Indeks Harga Saham Gabungan (IDX)

Muhammad Affarel, Fikri Ikhsan Ramadhan, Nugroho Aldi Prayoga — Wed, 31 Dec 2025 00:00:00 +0700

Penelitian ini mengevaluasi kinerja algoritma Reinforcement Learning Q-Learning dan SARSA dalam menghasilkan strategi trading otomatis pada lima saham Bursa Efek Indonesia (BBCA, BBRI, TLKM, UNVR, dan ASII) melalui simulasi 1.000 episode. Analisis dilakukan berdasarkan pola reward, equity curve, dan statistik performa akhir untuk mengukur efektivitas pembelajaran pada kondisi pasar yang berbeda. Hasil penelitian menunjukkan bahwa Q-Learning lebih unggul pada saham dengan momentum harga kuat karena sifat eksplorasinya yang lebih agresif, sedangkan SARSA memberikan performa yang lebih stabil pada pasar dengan volatilitas tinggi karena pendekatannya yang konservatif dan on-policy. Secara keseluruhan, kedua metode tidak menunjukkan dominasi absolut, namun menawarkan keunggulan berbeda sesuai karakteristik saham dan profil risiko strategi. Temuan ini menegaskan potensi RL untuk pengembangan algorithmic trading di pasar Indonesia dan membuka peluang eksplorasi model lanjutan yang lebih adaptif

Perancangan Aplikasi Mobile Pengelolaan Data Mahasiswa Berbasis Kodular Dengan Integrasi Airtable

Wed, 31 Dec 2025 00:00:00 +0700

Penelitian ini bertujuan untuk merancang dan mengimplementasikan aplikasi mobile pengelolaan data mahasiswa menggunakan platform no-code Kodular yang terintegrasi dengan Airtable sebagai cloud database. Aplikasi ini dikembangkan untuk mempermudah proses administrasi data mahasiswa seperti input, edit, hapus, tampilan (view), dan pembaruan (refresh) data secara daring. Pengembangan dilakukan dengan menggunakan metode Waterfall yang terdiri dari lima tahapan, yaitu analisis kebutuhan, perancangan sistem, implementasi, pengujian, dan pemeliharaan. Implementasi dilakukan dengan membangun satu layar utama (single screen application) yang dilengkapi form input NIM, nama, jurusan, dan angkatan serta lima tombol utama yang terhubung ke Airtable melalui API. Pengujian menggunakan metode black-box menunjukkan bahwa seluruh fungsi berjalan sesuai spesifikasi dan aplikasi dapat melakukan sinkronisasi data secara real-time. Hasil penelitian menunjukkan bahwa aplikasi ini layak digunakan sebagai solusi digital untuk pengelolaan data akademik sederhana dan dapat menjadi referensi pengembangan aplikasi serupa di bidang pendidikan.

Analisis Sentimen Ulasan Google Play Store: Studi Komparatif Algoritma SVM, Naïve Bayes, dan Logistic Regression

Singgih Jatmiko, Charles Dometian — Sat, 10 Jan 2026 00:00:00 +0700

This research aims to compare Support Vector Machine (SVM), Naïve Bayes, and Logistic Regression methods in sentiment analysis of app reviews on Google Play Store to identify the best method based on accuracy, precision, recall, and F1-Score using 2000 GoPay and LinkAja reviews from Google Play Store respectively. The methodology consists of six stages, namely, data collection, labeling method evaluation, preprocessing evaluation, SMOTE testing to overcome imbalanced data, hyperparameter tuning optimization, and consistency validation with a combination of TF-IDF and three classification methods. The data were split using an 80:20 ratio, with 80% of the data used for training and 20% for testing. Experimental results show SVM gives the best performance with 93% accuracy, 92% precision, 93% recall, and 92% F1-Score on the GoPay dataset due to its ability to find the optimal hyperplane, followed by Logistic Regression with 92% accuracy and the third Naïve Bayes despite identical accuracy but showing greater bias towards the majority class. Validation using the LinkAja dataset proves SVM still maintains the best performance with 95% accuracy, so the research concludes SVM is the best method for sentiment analysis of app reviews on the Google Play Store which is proven to provide optimal and consistent performance

Model Rekomendasi Destinasi Wisata Kreatif di Indonesia Berdasarkan Data Cuaca dan Ulasan Wisatawan

Kharisma, Irmma Dwijayanti, Ulfi Saidata Aesyi, Alfirna Rizqi Lahitani — Tue, 06 Jan 2026 00:00:00 +0700

Indonesia holds vast potential for creative tourism through its rich cultural heritage, natural beauty, and local creativity. However, travelers still face challenges in planning optimal trips due to the lack of context-aware and real-time recommendation systems. In practice, tourists often rely on Google Maps reviews, which are unorganized thematically, and there is limited integration with weather conditions—an important factor that significantly impacts travel experiences, particularly for nature-based destinations. This study aims to develop a recommendation model for creative tourism destinations in Indonesia by integrating two key aspects: sentiment analysis of Google Maps reviews and real-time weather data. The research utilizes tourist reviews from Google Maps alongside up-to-date weather information from destinations across Indonesia. The reviews are analyzed using the Support Vector Machine (SVM) algorithm to classify sentiments as positive or negative. These sentiment results are then combined with real-time weather data to build a Content-Based Filtering (CBF) recommendation system capable of providing more relevant and adaptive suggestions. The study successfully produced a recommendation system model with a testing accuracy of 90%.

Pendekatan Convolutional Neural Network dalam Mendeteksi Kemiringan Tulisan Tangan Menggunakan Framework YOLO

Anna Nurlita, muhammad haviz irfani, Zaid Romegar Mair — Sat, 10 Jan 2026 00:00:00 +0700

Despite the continuous advancement of digital technology, handwriting still plays an important role, especially in the field of education as a means of evaluating students’ writing skills. However, manual handwriting assessment tends to be subjective and inconsistent, particularly in the aspect of slant, which can reflect the clarity, legibility, and personality of the writer. Therefore, an automated method capable of accurately and objectively detecting handwriting slant is required. This study aims to develop an automated system based on a Convolutional Neural Network (CNN) using the YOLOv5 framework to detect the handwriting slant of university students. The dataset consists of 680 handwriting images annotated into three categories: upright, left-slanted, and right-slanted. The training process was conducted through four main experiments with variations in parameters such as batch size, epoch, and image size. The best model configuration was achieved with a batch size of 16, 150 epochs, and an image size of 640, resulting in an mAP@0.5 score of 0.894 and an F1-score of 0.84 on the training data. Evaluation on the training data showed that the model successfully classified left-slanted handwriting with 97% accuracy, right-slanted with 95%, and upright with 84%. On the test data, the model also demonstrated good performance with an average mAP@0.5 of 0.59, recall of 0.835, and classification accuracies of 100% for left-slanted, 93% for right-slanted, and 57% for upright handwriting. This study demonstrates that the CNN approach using YOLOv5 is effective for handwriting slant detection and has great potential for application in other related fields

Analisis Sentimen Ulasan Pemain Genshin Impact Menggunakan Kombinasi TF-IDF, Lexicon, dan Support Vector Machine

Danang Arbian Sulistyo, Mochammad Fiqi Fahrudillah — Fri, 09 Jan 2026 00:00:00 +0700

The rapid growth of the digital gaming industry in Indonesia has been accompanied by a significant increase in user-generated reviews on distribution platforms such as Google Play Store. This condition necessitates automated methods capable of efficiently interpreting player perceptions on a scale. This study conducts sentiment analysis on player reviews of Genshin Impact by developing a seven-stage analytical pipeline consisting of data preparation, lexicon-based labeling, TF-IDF feature extraction, Support Vector Machine (SVM) training, multi-metric evaluation, rule-based post-processing, and automated summarization using a Large Language Model. A total of 40,000 reviews from 2023 until 2025 were collected through web scraping and processed through text cleaning, slang normalization, tokenization, stopword removal, and stemming. Initial labels were generated using an updated domain-specific sentiment lexicon and subsequently refined through a rule-patch mechanism that handles negation, contrastive expressions, and domain-specific technical cues such as lag, bug, and crash. The SVM model was trained using a TF-IDF configuration (1–3 grams) and evaluated across 10 runs with different random seeds, producing an average accuracy of 0.945, a macro-F1 of 0.900, and stable performance across iterations. Visualization of sentiment distribution and WordClouds highlights prominent thematic patterns within each class, while automated summarization using IBM Granite provides qualitative insights into player appreciation of visual and character design, alongside complaints related to performance issues and the game’s gacha system. Overall, the integration of statistical, rule-based, and LLM-driven approaches demonstrates an effective and contextually robust framework for sentiment analysis in game analytics

Implementasi YOLOv8 untuk Deteksi Pelat Nomor dan Validasi Pajak Kendaraan

Muhammad Wesya Pralega, Witta Listiya Ningrum — Thu, 15 Jan 2026 00:00:00 +0700

The increase in the number of vehicles in Indonesia requires a more efficient administrative system for vehicle validity validation. Manual verification processes such as checking vehicle registration certificates and license plates by officers in the field are considered ineffective, prone to error, and time-consuming, especially when dealing with high volumes of vehicles. This study aims to develop a computer vision-based automated system capable of detecting vehicle license plates and independently validating tax status. The method used is the CRISP-DM method, which includes understanding requirements, data processing, modeling, evaluation, and implementation. The model used is YOLOv8 to detect the license plate area, and EasyOCR is used for alphanumeric character recognition. The research dataset consists of 587 secondary images and 15 primary images. The secondary data was divided into 70% training data, 20% validation data, and 10% test data. The YOLOv8 model was trained using the best combination of hyperparameters, namely 200 epochs, batch size 16, and learning rate 0.01, which produced a box loss value of 0.38. The tax status validation process is divided into four categories: active, expired, invalid, and no tax information available. Thus, this research can contribute to the development of an effective vehicle tax validation automation system that has the potential to be implemented in public administration services.

Analisis Penerapan Metode WASPAS untuk Penentuan Pola Belajar Mahasiswa Berdasarkan Gaya Belajar

Ria Ester, Dian Tri Yuniarti, Putri Eka Valentina, Faris Maulana Kusumah Putra — Sat, 10 Jan 2026 00:00:00 +0700

Higher education in the digital era requires learning approaches that are able to adapt to individual student characteristics, including differences in learning styles. This study aims to develop a model for assessing students’ learning patterns and to provide more personalized learning recommendations using the Weighted Aggregated Sum Product Assessment (WASPAS) method. The data used are secondary data obtained from 1,000 students with seven learning criteria, namely academic score, course participation, attendance rate, physical activity, emotional engagement, device usage, and feedback score. The WASPAS method is applied through two main stages, namely the calculation of the Weighted Sum Model (WSM) and the Weighted Product Model (WPM), which are then aggregated to produce a composite WASPAS score for each student. Manual calculations are demonstrated using five student samples, while computations for the entire dataset are performed using Python in the Jupyter Notebook environment. The results show that students’ WASPAS scores range from 0.2815 to 0.9914 with a distribution that tends to be normal. Most students fall into the “fair” to “very good” learning pattern categories, while a small proportion are classified as “very high” and “requiring special attention.” Analysis based on visual, auditory, and kinesthetic learning styles indicates differences in average WASPAS scores across groups, supporting the effectiveness of the WASPAS method in integrating multiple learning criteria simultaneously. These findings demonstrate that WASPAS can be used as a decision support tool to map student learning profiles and assist in designing more adaptive, targeted, and personalized learning strategies in higher education

Analisis Perbandingan Metode DBSCAN dan Mean Shift Dalam Mengelompokkan Data IPM Kabupaten/Kota se-Indonesia

Tue, 06 Jan 2026 00:00:00 +0700

This study extends previous research that clustered the 2019 Human Development Index (HDI) data of regencies and cities in Indonesia using K-Means, K-Medoids, and Agglomerative Hierarchical Clustering (AHC). HDI is an important indicator for describing the level of regional development; therefore, clustering analysis of HDI data is needed to support more targeted development policy formulation. However, these conventional clustering methods have limitations, including the requirement to predefine the number of clusters and their limited ability to handle noise. Therefore, this study applies and compares two density-based clustering algorithms, namely DBSCAN (Density-Based Spatial Clustering of Applications with Noise) and Mean Shift, which are capable of forming clusters automatically without specifying the number of clusters in advance and can effectively handle noise. The determination of optimal parameters for each method is conducted using the Sw/Sb Ratio metric, which measures the ratio between within-cluster and between-cluster standard deviations. The results show that Mean Shift with an optimal bandwidth parameter of 1 achieves an Sw/Sb Ratio value of 0.3609, which is better than DBSCAN with a value of 0.3739, and also outperforms the clustering methods used in previous studies, which produced a value of 0.51. These findings indicate that density-based clustering algorithms, particularly Mean Shift, provide more representative clustering results for HDI data and may serve as a more effective alternative method for analyzing human development data in Indonesia.

Analisis Perbandingan Tools Forensic Pada Aplikasi Facebook Messenger Menggunakan Metode National Institute of Standards Technology (NIST)

Desti Mualfah, Febri Israndi, Rizdqi Akbar Ramadhan — Wed, 31 Dec 2025 00:00:00 +0700

The rapid growth of internet and social media usage has led to an increase in cybercrime. This study aims to analyze digital forensic processes using the National Institute of Standards and Technology (NIST) method on Facebook to combat cybercrime. Two forensic tools, Magnet Axiom and MOBILedit, were utilized to collect and analyze digital evidence. The research employed a case study and analysis using the NIST method. Results show that MOBILedit excels in data collection, particularly in retrieving images, whereas Magnet Axiom boasts superior data analysis and integration capabilities. Consequently, MOBILedit is recommended for Android digital forensic applications.

Analisis Sentimen Calon Gubernur Jawa Tengah 2024 Menggunakan Metode Naïve Bayes

Martin Nuh Hanan, Hamid Muhammad Jumasa, Ike Yunia Pasa — Wed, 31 Dec 2025 00:00:00 +0700

Social media platform X (formerly Twitter) has become a public space where people can freely express their opinions, including in the context of regional elections. These opinions can be processed into useful information for decision-makers, especially in political contexts. This study aims to analyze public sentiment toward the candidates for Governor of Central Java for the 2024–2029 period using the Naïve Bayes method.

The data was collected through a crawling process on X using Tweet-Harvest and relevant keywords. The raw data then underwent preprocessing, including cleaning, case folding, normalization, stopword removal, tokenization, and stemming. Sentiment labeling was performed automatically using the TextBlob library, which classified tweets into positive, negative, or neutral categories. Naïve Bayes was chosen for its effectiveness and efficiency in text classification tasks.

The results showed model accuracy of 90.28% for Andika Perkasa and 84.51% for Ahmad Luthfi, using a 90:10 training-to-testing data ratio. Out of 452 total tweets, Andika Perkasa received 350 positive sentiments, slightly more than Ahmad Luthfi, who received 336. These findings indicate that public perception toward both candidates is generally positive, with a slight edge for Andika Perkasa.

Implementasi Convolutional Neural Network Pada Deteksi Penyakit Retina Menggunakan Citra Fundus Mata

Reimon Aldin — Tue, 20 Jan 2026 00:00:00 +0700

This study aims to develop an efficient system for detecting retinal diseases from fundus eye images by applying Deep Learning with Convolutional Neural Networks (CNN). The proposed model is designed to assist ophthalmologists in making accurate and timely diagnoses, enabling appropriate treatment and improving patient outcomes. The research emphasizes the role of CNN-based Deep Learning as a reliable method for classifying retinal disorders. A quantitative approach was employed, utilizing numerical and descriptive data such as images, observations, and secondary sources. The research procedure covered several stages: image preprocessing, CNN model design, training, validation, evaluation, and system testing. The experimental results demonstrated that the developed system achieved an accuracy of 97%. Evaluation metrics confirmed high performance with classification results as follows: Myopia (precision 1.00, recall 1.00, f1-score 1.00), Cataract (precision 0.88, recall 1.00, f1-score 0.93), Diabetic Retinopathy (precision 1.00, recall 1.00, f1-score 1.00), and Glaucoma (precision 1.00, recall 0.95, f1-score 0.97). These findings show that the CNN architecture with VGG16 demonstrates excellent capability in detecting and classifying retinal diseases using fundus images. Therefore, the model can be recommended as a practical tool for early detection of retinal disorders, particularly within the context of healthcare services in Kendari City, Southeast Sulawesi.

Penerapan Empirical-Bayes pada Sistem Peringkat Produk E-Commerce

Chandra Pratama, Fahri Ramadhan, Ghibran Arrazi Satria, Aji Setiawan — Fri, 09 Jan 2026 00:00:00 +0700

This study examines the application of Empirical Bayes (EB) smoothing for product ranking in e-commerce platforms characterized by sparse sales signals and highly skewed transaction distributions. Under these conditions, top lists tend to fluctuate when rankings rely solely on raw cumulative sales, particularly for long-tail products; therefore, a method that balances population-level information with item-level evidence is required to produce more consistent top-k rankings. The method models purchase counts using a Gamma–Poisson framework, where a global prior is estimated from the overall data and item-level posteriors are updated so that the posterior mean serves as a smoothed popularity score. Experiments are conducted on real product catalogs (smartphones and laptops) augmented with a 12-week sales simulation featuring mild seasonality and promotional noise, and EB is compared against a naive baseline that ranks items by raw cumulative units sold under a rolling, week-by-week evaluation. Results show that EB improves NDCG@5 and NDCG@10 while reducing week-to-week Top-10 churn relative to the baseline, with the most notable gains observed for low-signal and long-tail items because shrinkage dampens extreme rank swings caused by sparse observations. Overall, EB smoothing is effective in stabilizing top-k product rankings for listing interfaces and administrative dashboards, and it can be extended through time-decayed priors and the incorporation of contextual features such as price and category to further improve ranking accuracy

Kolaborasi Algoritma K-Nearest Neighbor Dan Gradient Boosting Untuk Klasifikasi Diabetes Melitus Tipe 2

Wed, 31 Dec 2025 00:00:00 +0700

Diabetes Melitus Tipe 2 (DMT2) telah menjadi salah satu tantangan kesehatan masyarakat terbesar di Indonesia, dengan prevalensi yang terus meningkat dan sebagian besar kasus tidak terdiagnosis. Deteksi dini menjadi kunci untuk mencegah komplikasi serius. Penelitian ini bertujuan untuk mengembangkan dan mengevaluasi model klasifikasi berbasis machine learning untuk prediksi DMT2. Tiga pendekatan dieksplorasi: algoritma K-Nearest Neighbor (KNN), Gradient Boosting, dan model KNN + Gradient Boosting yang mengintegrasikan keduanya melalui arsitektur stacking ensemble. Kinerja diukur menggunakan metrik akurasi, presisi, recall, dan F1-score. Hasil penelitian menunjukkan bahwa model ensemble secara signifikan mengungguli model tunggal. Model KNN mencapai akurasi 90.92% namun dengan presisi yang rendah untuk kelas diabetik (0.48). Model Gradient Boosting menunjukkan peningkatan signifikan dengan akurasi 95.50% dan presisi 0.72. Model KNN + Gradient Boosting menunjukkan kinerja terbaik dengan akurasi keseluruhan 96.17% dan presisi tertinggi untuk kelas diabetik (0.81), yang secara efektif mengurangi tingkat alarm palsu. Temuan ini mengindikasikan bahwa model KNN + Gradient Boosting memiliki potensi besar sebagai alat bantu pendukung keputusan klinis yang andal untuk identifikasi dini individu berisiko tinggi DMT2.

Analisis Sentimen Media Sosial X Terhadap Kebijakan Presiden Republik Indonesia Prabowo Subianto

Ina Najiyah, Miftahul Rizal — Wed, 31 Dec 2025 00:00:00 +0700

This study aims to identify and measure the tendency of public sentiment towards the implementation of the policies of the President of the Republic of Indonesia, Prabowo Subianto. The methodology used is text mining-based sentiment analysis, utilizing a data corpus taken from the social media platform X. This study adopts the SEMMA (Sample, Explore, Modify, Model, Assess) workflow as a procedural framework. Data retrieval is carried out automatically using crawling techniques. Next, the data goes through a comprehensive text pre-processing stage, including cleaning, case folding, normalization, convert negation, tokenizing, stopword removal, stemming. Sentiment polarity is determined automatically through a lexicon-based approach, implemented with the VADER (Valence Aware Dictionary for Sentiment Reasoning) algorithm. The modeling phase uses two machine learning classification algorithms, namely Naïve Bayes and Support Vector Machine (SVM). Performance testing is carried out on three different training and testing data distribution schemes (90:10, 80:20, and 70:30). The evaluation findings show that the Naïve Bayes algorithm achieved the highest accuracy rate of 81.25% at a ratio of 80:20. Meanwhile, SVM consistently recorded superior accuracy, reaching a maximum value of 92.60% at a ratio of 90:10. Based on a comprehensive assessment of performance metrics (accuracy, precision, recall, and f1-score), the Support Vector Machine (SVM) algorithm was proven to provide significantly superior performance compared to Naïve Bayes in this sentiment classification task

Klasifikasi Mahasiswa Calon Penerima Beasiswa KIP Menggunakan Algoritma Naive Bayes di Universitas Tomakaka Mamuju

Hidayat, Musliadi KH, Indar Kusmanto, Munawirah Kadir, Krsitian — Wed, 31 Dec 2025 00:00:00 +0700

Education is a fundamental aspect of national development that demands equal access to quality education for all. The Smart Indonesia Card (KIP) program is a government initiative aimed at supporting education for underprivileged communities. Tomakaka University, Mamuju, as one of the universities in West Sulawesi, plays an active role in distributing KIP scholarships to students who meet certain criteria. However, the selection process for prospective scholarship recipients has been carried out manually, which may lead to inefficiencies and inaccurate targeting. This study aims to apply the Naïve Bayes algorithm to classify prospective KIP scholarship recipients to make the selection process more objective, fast, and accurate. The research method uses a data mining approach with stages of data preprocessing, dividing training and test data, model training, and testing using the Python programming language on the Google Colab platform. The dataset used is 171 student data, with a division of 75% training data and 25% test data. The test results showed that the Naïve Bayes model achieved an accuracy of 95.35%, with a precision of 97%, a recall of 97%, and a loss of 4.65%, indicating excellent classification performance. Thus, this research contributes to improving administrative efficiency and targeting of KIP scholarship distribution at Tomakaka University, Mamuju.

Prediksi Dropout Mahasiswa: Early-Warning Berbasis Enrollment dengan Machine Learning

Wed, 31 Dec 2025 00:00:00 +0700

Dropout among university students remains a major challenge in higher education because it affects study continuity, institutional performance, and the efficiency of academic service planning. This study develops a machine learning–based Early Warning System (EWS) that leverages data available at enrollment and is updated after the first semester. Using the public dataset “Predict Students’ Dropout and Academic Success” (n = 4,424), the original three-class outcome (Dropout, Enrolled, Graduate) is simplified into a binary target, with dropout treated as the positive class. Two feature scenarios are evaluated: (1) enrollment-only for pre-entry screening and (2) enrollment plus first-semester indicators to update risk scores. Three models are compared: class-balanced Logistic Regression, class-balanced Random Forest, and Gradient Boosting. Model performance is assessed using accuracy, precision/recall/F1score for the dropout class, balanced accuracy, and ROC-AUC. Under the enrollment-only setting, Logistic Regression achieves the best early-warning performance (recall = 0.697; F1 score = 0.651). After incorporating first-semester features, performance improves (recall = 0.792; F1score = 0.779). Beyond model comparison, this study adds an operational perspective through confusion-matrix simulation and probability-threshold analysis to balance missed at-risk cases and follow-up workload.

Klasterisasi Topik Khotbah Pendeta Di GBI MPI Palembang Dengan Metode DBSCAN

Kristian Fernando, Hafiz Irsyad — Wed, 31 Dec 2025 00:00:00 +0700

Comprehensive evaluation of the teaching curriculum proportion at GBI Rayon 15 Musi Palem Indah (MPI) Palembang is a fundamental element in ensuring the doctrinal health of the congregation. However, the current evaluation process is inefficient due to reliance on manual mapping of ever-growing sermon archives. This conventional method carries a high risk of subjectivity bias, making it difficult for church leadership to objectively observe teaching theme trends. This study addresses this issue by developing an automated document clustering system based on Text Mining to process 406 sermon summary documents from the 2023-2025 period. The methodology includes preprocessing, Term Frequency-Inverse Document Frequency (TF-IDF) weighting to highlight distinctive theological terms, and the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm. DBSCAN was specifically selected for its superiority in handling data with varying densities and its ability to isolate outliers without requiring a static cluster count parameter. Test results indicate an optimal configuration at Epsilon 0.3 and MinPts 3, yielding very high internal validity with a Silhouette Coefficient of 0.8888 and forming 32 core topic clusters. Significant findings reveal a high noise ratio (71%), which effectively separates incidental topics, such as holiday celebrations, from regular material. Practically, these results serve as an early warning system mechanism for the church to detect doctrinal imbalances or material gaps, providing a strategic data-driven foundation for holistic curriculum improvement.

Implementasi Otomatisasi Data Lifecycle Management (DLM) untuk Peningkatan Skalabilitas dan Keandalan Sistem Informasi Pemesanan Kelas

Sybil Auzi, Atta Zulfahrizan, Dedi Kiswanto, Josua Tampubolon — Wed, 31 Dec 2025 00:00:00 +0700

The rapid growth of historical data in the class booking information system can significantly degrade performance, impacting system scalability and reliability. This research addresses the issue by designing and implementing an automated Data Lifecycle Management (DLM) framework. The primary objectives are: (1) to develop a functional automated DLM prototype using the Laravel framework and its task scheduler, and (2) to analyze how this implementation enhances system scalability and reliability. This study adopts a system implementation method by applying the seven stages of DLM, supported by a Dual Connection database architecture that separates operational data (Hot Storage) from historical archives (Cold Storage). The results demonstrate the successful implementation of all DLM stages, from data creation to automated deletion. The system automatically archives weekly transactional data and permanently deletes them after a retention period of one semester plus a 30-day grace period. Furthermore, a secure public API was developed to facilitate data sharing for academic purposes. The implementation of automated DLM proves effective in managing data lifecycle, reducing the burden on the primary database, and maintaining system performance, thereby ensuring better scalability and reliability.