Bulletin of Information Technology (BIT)

Implementing Mobile-based AI in Household Waste Type and Condition Classification

Suwarno Suwarno — 2026-03-26

Urbanization and population growth have significantly increased waste generation, creating challenges for effective waste management and recycling. Improper waste sorting and management often results to unrecyclable waste contaminating recycling streams or recyclable waste ending up in landfill. This research presents a mobile-based waste classification application that integrates YOLOv11n for real-time object detection, and uses TensorFlow Lite with a Flutter-based user interface. The model was trained on a dataset of 4,410 images, which combines self-gathered images and images from Kaggle dataset. The images are then augmented to 10,936 images covering 23 waste classes, including organic, inorganic, hazardous, and residual types, with their recyclability conditions. The application allows users to detect objects using their phone camera, to identify their classification and condition, as well as receive actionable 3R (Reduce, Reuse, Recycle) recommendations. Evaluation results show a precision of 0.5963, recall of 0.60563, mAP@0.5 of 0.62246, and mAP@0.5:0.95 of 0.5279, indicating decent classification despite challenges posed by visually similar objects and variable backgrounds. Overall, the system demonstrates the feasibility of deploying a lightweight AI model on mobile devices in hopes of supporting proper waste segregation, increase user awareness, and potentially reduce contamination in recycling streams through practical waste classification.

Performance Analysis of XGBoost in Handling Missing Data on the Telco Customer Churn Dataset

muhammad riki atsauri — 2026-03-26

This study analyzes the performance of Extreme Gradient Boosting (XGBoost) algorithm in handling missing data for telecommunications customer churn prediction. The research objective is to compare the effectiveness of various missing data imputation techniques (mean, k-NN, and MICE) on XGBoost performance using the IBM Telco Customer Churn dataset. The research methodology includes data preprocessing, implementation of imputation techniques, XGBoost model training, and evaluation using accuracy, precision, recall, and F1-score metrics. The results show that MICE imputation technique provides the best performance improvement with 81.24% accuracy, 69.80% precision, 58.40% recall, and 63.60% F1-score, compared to XGBoost without imputation achieving 79.43% accuracy. These findings demonstrate that explicit missing data handling can enhance XGBoost's predictive capability in identifying potential churning customers. The practical implications of this research provide guidance for telecommunications industry in optimizing customer retention strategies through more accurate churn prediction

Pemodelan Biaya Sewa pada Data Pendidikan Internasional Menggunakan Pendekatan Machine Learning dan CRISP-DM

Arif Nababan — 2026-03-26

Advances in machine learning drive its application in analyzing complex educational data. In international education, housing rent (Rent_USD) is a critical cost-of-living component showing significant variation across regions. These variations are influenced by geography, local economics, and educational environments, requiring systematic data modeling. This study aims to model Rent_USD using the CRISP-DM framework: Business Understanding, Data Understanding, Data Preparation, Modeling, and Evaluation. Three algorithms were employed: Decision Tree as the baseline, Random Forest as a comparison, and XGBoost as the primary model. To enhance performance, hyperparameter tuning was conducted via GridSearchCV. Model evaluation utilized Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R²). The experimental results demonstrate that the XGBoost algorithm delivers the most superior performance, achieving the lowest RMSE of 93.27 USD and an R²of 0.96. This performance outperforms Random Forest (RMSE: 114.87, R²: 0.94) and Decision Tree (RMSE: 157.16, R²: 0.89). Furthermore, feature importance analysis revealed crucial findings: the Living Cost Index and Tuition Fee are the most dominant factors influencing Rent_USD variations, contributing 58.32% and 32.94% respectively. This research provides an empirical overview of machine learning applications in modeling international education costs and serves as a vital reference for future studies regarding educational data management and predictive analytics in global student mobility.

Analyzing COVID-19's Educational Impact in Indonesia: K-Means and Self-Organizing Map Approach

Ika Nur Laily Fitriana — 2026-03-26

The COVID-19 pandemic has affected the education sector. This research aimed to investigate the impact of COVID-19 on the education sector in Indonesia, especially on school participation indicators, using cluster analysis. We used fifteen factors related to the involvement indicators of students in elementary, junior secondary, and senior secondary education. The comparison of factors between 2019 and 2020 related to the effects of COVID-19, which began to proliferate in Indonesia in March 2020. Consequently, comparing those periods yields insights into the timeframe before and after the spread of COVID-19. To assess the pandemic's influence on the education sector, we performed an inferential statistical analysis using a nonparametric location test to identify significant changes between variables in 2019 and 2020. Subsequently, we performed cluster analysis using K-Means and Self-Organizing Map (SOM) approaches. The optimal cluster obtained for K-Means and SOM is three clusters. The results indicate that SOM and K-Means exhibit similar performances. Changes in cluster members in 2019 and 2020 indicate an enormous impact due to COVID-19. Cluster 3, which consists of DKI Jakarta, West Java, Central Java, East Java, and North Sumatra, is most affected by the pandemic from the educational sector.

Implementasi Support Vector Machine (SVM) Untuk Deteksi Serangan Jaringan Pada Sistem Keamanan Jaringan Kampus

Mochammad Darip — 2026-03-26

Network security in campus environments faces increasingly complex challenges due to the rapid growth of internet usage, digital academic systems, and the large number of devices connected to the network. One of the main problems is the limitation of conventional security systems in detecting new or anomalous network attacks. Traditional systems generally rely on predefined attack signatures, making them ineffective against previously unknown attacks. Therefore, this study proposes a solution by implementing the Support Vector Machine (SVM) method for automatic network attack detection. The research method includes the collection of campus network traffic data, data preprocessing stages such as data cleaning, normalization, and feature selection, SVM model training, and performance evaluation using confusion matrix and ROC curve. The results show that the SVM model is able to classify normal traffic and attack traffic with very high accuracy. These findings indicate that SVM is an effective method for intrusion detection and can significantly enhance campus network security in an adaptive and efficient manner.

Model Integrasi Machine Learning dan Decision Support System dalam Pemetaan Potensi UMKM Kabupaten Polewali Mandar

Basri — 2026-03-26

Abstrak- Sektor Usaha Mikro, Kecil, dan Menengah (UMKM) memegang peranan krusial dalam struktur ekonomi daerah, namun di Kabupaten Polewali Mandar, pengembangan sektor ini masih menghadapi kendala signifikan karena potensi wilayah yang belum terpetakan secara komprehensif berbasis data digital. Meskipun Produk Domestik Regional Bruto (PDRB) daerah didominasi oleh sektor pertanian sebesar 46,58% dan pertumbuhan sektor perdagangan mencapai 7,93% pada tahun 2024, distribusi sumber daya dan kebijakan pendukung UMKM seringkali belum tepat sasaran akibat ketiadaan model klasifikasi wilayah yang objektif. Penelitian ini bertujuan untuk mengembangkan model integrasi antara Machine Learning (ML) dan Decision Support System (DSS) guna memetakan potensi UMKM di 16 kecamatan Kabupaten Polewali Mandar. Metodologi yang digunakan adalah algoritma K-Means Clustering untuk pengelompokkan wilayah dan metode pembobotan Analytic Hierarchy Process (AHP) untuk menentukan prioritas kriteria. Data penelitian bersumber dari Badan Pusat Statistik Kabupaten Polewali Mandar Tahun 2025, mencakup variabel PDRB sektoral, statistik tenaga kerja, dan akses kredit usaha.Hasil penelitian menunjukkan terbentuknya tiga cluster wilayah utama, yaitu potensi tinggi (pusat pertumbuhan), potensi sedang (wilayah berkembang), dan potensi rendah (wilayah tertinggal). Evaluasi model menggunakan Silhouette Score menghasilkan nilai 0,62, yang menunjukkan bahwa pengelompokkan wilayah memiliki struktur cluster yang cukup kuat dan baik. Implementasi model ini memberikan rekomendasi strategis bagi pemerintah daerah dalam mengalokasikan bantuan dan infrastruktur pendukung UMKM secara presisi sesuai karakteristik ekonomi masing-masing kecamatan untuk mendukung peningkatan Indeks Pembangunan Manusia (IPM) yang kini berada pada angka 70,71.

Analisis Komparatif Algoritma Klasifikasi untuk Prediksi Kelulusan Tepat Waktu Mahasiswa

Hariati Husain — 2026-03-28

- Timely student graduation is an important indicator in assessing the quality of higher education management. However, not all students are able to complete their studies within the prescribed study period, making it necessary to implement data-driven predictive approaches to identify students at risk of delayed graduation. This study aims to compare the performance of the Decision Tree and Naïve Bayes algorithms in classifying timely student graduation based on academic data. The dataset consists of alumni records from the Informatics Engineering Study Program for the 2015–2016 cohorts, totaling 610 valid records after data cleaning and attribute selection. Predictor variables include gender, class type, and Semester Grade Point Index (IPS) from semester 1 to semester 5, while the target variable is graduation status. Model evaluation was conducted using an 80% training and 20% testing split, and performance was measured through a confusion matrix to obtain accuracy, precision, and recall values. The results show that the Decision Tree achieved an accuracy of 69.54%, while Naïve Bayes achieved 68.38%. The 1.16% difference indicates that the Decision Tree performs slightly better for this dataset. These findings suggest that early semester academic performance significantly contributes to predicting timely graduation and can support data-driven academic decision-making.

Model Ensemble Fusion–Stacking untuk Klasifikasi Varietas Salak Berbasis Deep Feature

Bunga Intan — 2026-03-28

Visual classification of salak (snake fruit) varieties remains challenging due to similarities in texture, color, and morphological characteristics across classes. Manual identification is prone to subjectivity and inconsistency in determining varieties. This study proposes an ensemble model based on fusion and stacking applied to deep learning feature extraction in order to improve the accuracy of salak variety classification. Image features are extracted using two pre-trained Convolutional Neural Network architectures, namely VGG16 and ResNet50, as deep feature extractors. The resulting feature representations are subsequently classified using K-Nearest Neighbor (KNN) and Support Vector Machine (SVM) algorithms. The output probabilities of both classifiers are then combined through a stacking ensemble approach with Logistic Regression as the meta-learner. The dataset consists of 584 images distributed across four salak varieties. Model performance is evaluated using accuracy, precision, recall, and F1-score metrics. Experimental results indicate that the proposed fusion–stacking approach achieves an accuracy of 95%, outperforming single CNN-based models and conventional classification methods. These findings demonstrate that the integration of deep feature extraction and ensemble learning effectively enhances the discriminative capability of the model in agricultural image classification.

Prototipe Smart Traffic Light (STL) berdasar Panjang Antrian menggunakan Internet of Things (IoT)

Jihan Athira Ramadhani — 2026-03-28

As an effort to regulate traffic, a Traffic Instruction Control Device (APIL) in the form of a Traffic Light is used. Traffic Lights are installed at various types of road intersections and crossing facilities. The function of traffic lights is very important so that traffic lights must be controlled as easily and efficiently as possible to facilitate traffic flow at a road intersection. However, after the use of Traffic Lights, there is still congestion due to the direction of vehicle arrivals from each lane not being simultaneous. This results in long queues in one of the Traffic Light lanes which makes the queue even longer due to the ineffectiveness of the red light duration of the existing Traffic Light. Therefore, researchers want to design a tool using 8 IR sensors as queue detection in each lane, then combining it with an automatic Traffic Light system can optimize the red and green lights according to the existing queue length. The results of the experiment found that in quiet conditions or no sensors detecting vehicles, the green light time is 20 seconds. When the condition of congestion 1 or the front IR sensor detects a vehicle, the green light time is increased by 5 seconds to 25 seconds. When traffic jam condition 2 or both IR sensors detect a vehicle, the green light is extended by 10 seconds to 30 seconds.

Analisis Performa Support Vector Machine untuk Klasifikasi Risiko Kredit Nasabah pada Perbankan Daerah

Asep Sapaatullah — 2026-03-29

Credit risk assessment is a crucial component of the banking system because it directly relates to a financial institution's ability to manage potential losses due to non-performing loans. Banks often face difficulties in accurately classifying customer credit risk levels, especially when the data being analyzed is complex, nonlinear, and contains interacting variables. Conventional methods such as regression analysis often fail to capture hidden patterns in such data. Therefore, this study aims to apply the Support Vector Machine (SVM) algorithm as a solution to classify bank customers' credit risk levels based on attributes such as income, loan amount, length of employment, payment history, debt-to-income ratio, and asset ownership status. The research process begins with data collection and pre-processing, including data cleaning and normalization to ensure a uniform distribution of values. The data is then divided into training and test data with specific proportions. An SVM model is then applied using several kernel types, such as linear, polynomial, and radial basis function (RBF), to determine the best-performing kernel. Model evaluation is performed using accuracy, precision, recall, and F1-score metrics to measure classification performance. Test results show that the SVM model with the RBF kernel provided the best results, achieving an accuracy rate of over 90% and minimizing classification errors in the high-risk category. In conclusion, the application of the SVM algorithm has proven effective in classifying customer credit risk levels with high accuracy and stability, making it a reliable tool for banks in the creditworthiness analysis process and more accurate, data-driven strategic decision-making

Penerapan Metode Single Exponential Smoothing untuk Prediksi Penjualan Kacamata pada Optik Adis

Ade Rizka — 2026-04-04

Accurate inventory management is a crucial factor in supporting the sustainability of retail businesses, including those in the optical industry. Precision in determining the quantity of eyeglass inventory has a significant impact on operational efficiency and customer satisfaction. Optik Adis still faces challenges in estimating eyeglass stock requirements due to the manual process of recording and analyzing sales data. This situation has the potential to cause overstock conditions, leading to increased warehouse costs, or conversely, stock shortages that result in lost sales opportunities. Based on these issues, this study aims to develop an eyeglass sales prediction platform to support managerial decision-making in inventory management. The forecasting method used is Single Exponential Smoothing (SES) because it is suitable for sales data that fluctuates relatively stably. The system development process refers to the Waterfall methodology, which includes requirement analysis, system design, implementation, testing, and maintenance. Sales data from February to November 2025 were used as the basis for computation, with testing variations of the alpha parameter ranging from 0.1 to 0.9. The research findings reveal that an alpha value of 0.1 produces the best prediction performance with a Mean Absolute Percentage Error (MAPE) value of 19.06%, which falls into the “Good” category, and a Mean Absolute Deviation (MAD) value of 3.536 units. The application of the method in the system projects eyeglass sales in December 2025 at 17.50 units. Black Box testing shows that all system functions operate properly. Thus, the implementation of the method can help Optik Adis plan stock procurement more accurately, effectively, and efficiently.