Journal of System and Computer Engineering

Analysis Of A Textbook Recommendation System Based On Semantic Search And Full-Text Retrieval Based On Students' Learning Materials At The Labuang State Vocational High School Library

Muhammad Fadhil AY — 2026-10-30

School libraries play a crucial role in providing relevant learning resources for students. However, book search systems based on titles or simple keywords often fail to understand learning needs contextually. This study aims to develop a textbook recommendation system based on semantic search and full-text retrieval based on student learning materials at the Labuang State Vocational High School Library.

This study utilizes Natural Language Processing (NLP) techniques in the text pre-processing stage and uses the Term Frequency–Inverse Document Frequency (TF-IDF) method and sentence embedding to represent text. The level of similarity between teaching materials and book content is calculated using the cosine similarity method, then combined through a Hybrid Retrieval approach to produce a book recommendation relevance score. The research dataset consists of 80 textbooks that are processed into 21,556 text paragrap hs as the basis for the system's search.

System performance evaluation was conducted using Precision, Recall, F1-Score, and Mean Average Precision (MAP) metrics to measure the relevance and quality of document rankings generated by the system. The results showed that the Hybrid Retrieval approach was capable of generating relevant book recommendations based on student learning materials. The developed system is expected to assist students in finding books that meet their learning needs and improve the effectiveness of information searches in school libraries.

Development of a Mobile Application for Social Media Duration Tracking

Wakhid Yunendar — 2026-10-30

Self-report methods are known to inadequately represent actual social media use, limiting the accuracy of duration-comparison research among adolescents. This second-year study addressed this limitation by developing SociaLog, an Android mobile application that automatically logs students' social media duration and frequency of use. Using a Research and Development approach combined with the Prototyping software development model, the application was designed through five phases: requirements gathering, quick design, prototype building, user evaluation, and refinement. A preliminary needs analysis (n = 10) established four functional requirements, prioritizing automatic logging and graph-based reporting. The prototype was trial-tested on 100 vocational high school students from two schools in Makassar, Indonesia, over a one-week period. Logged data showed an average daily social media use of 7.46-8.63 hours, dominated by WhatsApp and TikTok. A 27-item usability questionnaire demonstrated strong item validity (r = 0.65-0.88) and excellent reliability (Cronbach's alpha = 0.980). Non-parametric tests indicated consistently high user acceptance across schools (Mann-Whitney U, p = 0.304) and evaluation dimensions (Kruskal-Wallis, p = 0.846). These findings indicate that SociaLog offers a valid, reliable, and well-accepted alternative to self-reported social media duration measurement among adolescents.

Optimizing a CNN–LSTM Model for Respiratory Disease Classification from Lung Sounds

Bahrum Herdania Bahar — 2026-10-30

Respiratory diseases remain a significant health concern, while accurate identification of respiratory conditions from lung sounds is challenging because acoustic characteristics can overlap among disease categories. This study aims to optimize a Convolutional Neural Network–Long Short-Term Memory (CNN–LSTM) model for multiclass respiratory disease classification from lung sound recordings while addressing class imbalance and limited subject diversity. The study used the Lung Dataset containing 336 recordings from 112 subjects. After excluding subjects with combined diagnoses and consolidating low-frequency categories, 324 recordings from 108 subjects representing seven classes were retained. The audio signals were normalized, standardized to 20 s, and transformed into 64-bin Mel-spectrograms with 160 time steps. CNN was employed to extract time-frequency features, while LSTM modeled temporal dependencies. Model optimization was further evaluated using subject-level 5-fold Group Cross-Validation, data augmentation, Focal Loss, and Class-Balanced Cross-Entropy. The baseline CNN–LSTM achieved an accuracy of 34.85%, precision of 24.61%, recall of 34.85%, and F1-score of 28.22% in the initial evaluation, with a macro-F1 of 14.07%. In the extended cross-validation experiment, the baseline CNN–LSTM achieved a mean accuracy of 35.87 ± 5.43% and macro-F1 of 10.26 ± 3.50%. Among the imbalance-handling strategies, Class-Balanced Cross-Entropy produced the highest macro-F1 of 12.16%, compared with 10.26% for the baseline, although its overall accuracy decreased to 27.52%. The results indicate that the principal challenge is not solely the model architecture but also severe class imbalance and the limited number of independent subjects. Therefore, subject-level validation and class-sensitive evaluation metrics are essential for obtaining a more reliable assessment of respiratory sound classification models. The optimized CNN–LSTM framework demonstrates potential for automated respiratory disease classification, although further validation using larger and more diverse datasets is required before clinical application.

Energy-Autonomous IoT Precision Irrigation With Soil Moisture Telemetry For Paddy

Muhammad Rijal — 2026-10-31

The adoption of IoT-based smart agriculture in Indonesia is still limited among farmers due to high costs, low digital literacy, and the lack of an easy-to-replicate system. This study develops and evaluates an IoT-based rice field irrigation monitoring and control system with a WiFi shooting and solar architecture. The system uses an ESP32 microcontroller, capacitive soil moisture sensor, DC water pump, 50 Wp solar panels, and TP-Link wireless transmitter for long-distance connectivity. The human-in-the-loop approach is implemented through a web dashboard that allows farmers to control irrigation manually while avoiding the risk of fully automated system failure. The performance of the system is evaluated through experimental testing including sensor validation against the Intelligent Soil Tester, network performance testing at distances of up to 1000 meters, and pump efficiency measurements. The results showed that the calibrated soil moisture sensor achieved 99.0% accuracy with an error of only 1%. The WiFi bridge maintains a stable connection up to a distance of 700 meters with an average response time of 2.08 seconds. The system reduces irrigation control time by 99.0% (from 15.5 minutes to 9 seconds per cycle), saving about 92 minutes each day. The pump produces an average discharge of 22.2 L/min with a power consumption of 180.1 W. This modular system offers an easy-to-replicate solution to bridge the gap between IoT research and the needs of farmers, while supporting the inclusive and sustainable digital transformation of Indonesia's food sector.

Vehicle Counting System on Ro-Ro Ferry Bengkalis Route Using YOLO26n and ByteTrack Algorithm

Muhamad Nasir — 2026-10-30

Vehicle transportation management on Ro-Ro ferries requires accurate information regarding the number and type of vehicles transported during the loading process. Conventional counting methods are still performed manually, which requires time and has the potential to produce human errors. This research proposes an automatic vehicle counting system on the Bengkalis Ro-Ro ferry route using YOLO26n object detection combined with ByteTrack multi-object tracking. The system processes video recordings from Ro-Ro ferry activities and detects three vehicle categories, namely cars, motorcycles, and trucks. The detection process uses YOLO26n with a confidence threshold of 0.25 and an image size of 640 pixels. ByteTrack is implemented to maintain vehicle identities and prevent duplicate counting when vehicles cross the counting line. The experimental results show that the proposed system successfully detected and counted 20 vehicles consisting of 5 cars, 10 motorcycles, and 5 trucks. The system achieved an average processing speed of 15 frames per second, indicating that the proposed approach can support real-time vehicle monitoring on Ro-Ro ferry transportation systems.

Fine Tuning Whisper Small for Low-Resource Makassar Automatic Speech Recognition

Hendra Sagita Putra — 2026-10-30

Automatic speech recognition (ASR) for regional languages is often constrained by limited paired speech–text data and by evaluation settings that can hide important differences between familiar and previously unseen transcript content. This study examines full-model fine-tuning of Whisper Small for Makassar ASR using a processed corpus of 6,119 utterances (6.3882 h). The corpus was partitioned into 4,894 training utterances, 612 validation utterances, and a locked 613-utterance test set. Because the same three speakers occur in all partitions, the evaluation is explicitly speaker-dependent. For a fair reviewer-driven baseline, the original zero-shot Whisper Small model and the fine-tuned checkpoint were re-evaluated on the identical test items with the same processor, 16-kHz audio preparation, Indonesian transcription setting, deterministic one-beam decoding, and a shared 225-token generation safety cap. Under this controlled protocol, overall WER decreased from 120.84% to 51.78%, while CER decreased from 62.00% to 16.19%. A post-hoc transcript-content audit identified 136 seen-content test utterances (22.19%) and 477 unseen-content utterances (77.81%). On unseen content, WER decreased from 125.88% to 60.91%, an absolute reduction of 64.97 percentage points. Error decomposition showed that substitutions remained the dominant error class after fine-tuning. The zero-shot system also showed repetitive long-generation behavior in targeted ceiling-hit cases. These findings support language- and corpus-specific adaptation of Whisper Small for Makassar, while also showing why aggregate error rates should be interpreted together with content-overlap, speaker-design, decoding, and reproducibility audits.

Preventive Voltage Drop System in Smart Farming Using Gradient Boosting

Riswandi Riswandi — 2026-10-30

Solar photovoltaic systems in smart farming frequently suffer from destructive voltage drops caused by weather fluctuations, endangering critical actuators like irrigation pumps and ventilation fans. To address this, this study presents a preventive control system integrating a Gradient Boosting Machine algorithm with real-time load management. The model utilizes weather, technical, and temporal features to forecast solar power generation. Training was conducted on 2,184 hours of historical data from April to June, while evaluation used a 744-hour test dataset from July collected at an 8,000-watt photovoltaic facility in South Sulawesi. Forecasted power output is fed directly into the load controller to prioritize actuator execution and enforce a maximum voltage drop limit below five percent. Experimental evaluation demonstrates superior predictive accuracy, achieving a Root Mean Square Error of 0.211 kilowatts and a Mean Absolute Error of 0.158 kilowatts. The findings confirm that proactive artificial intelligence load scheduling effectively prevents voltage degradation, offering a scalable, software-driven energy management solution for sustainable smart agriculture.

Random Survival Forest for Time-to-Maturity Prediction in Black Soldier Fly

ROZALINA AMRAN — 2026-10-30

Black soldier fly (BSF, Hermetia illucens) larvae are widely used to biologically convert organic waste into value-added biomass, but the time required to reach maturity varies with rearing temperature, humidity, and feed intake, which complicates harvest scheduling. This study developed and evaluated a Random Survival Forest (RSF) model to predict the time-to-transition to the adult (‘Dewasa’) stage of BSF under right-censoring, using temperature, humidity, feed consumption, larval count, and larval weight as covariates, on a monitoring dataset of 8,000 observations (28.1% events, 71.9% censored) over a 30-day rearing period. As groundwork, regression and classification benchmarks were also trained; Random Forest and Gradient Boosting explained most of the variance in total biomass (R² = 0.9998) and feed conversion ratio (R² = 0.929–0.938) and achieved near-perfect phase classification (accuracy 0.9999), results interpreted cautiously given likely redundancy between predictors and these outcomes. In contrast, the RSF model itself achieved a Harrell and Uno C-index of 0.490, an Integrated Brier Score of 0.213, and a mean time-dependent AUC of 0.489, indicating discrimination and calibration close to chance despite permutation importance detecting statistically distinguishable contributions from humidity, larval count, and temperature. These findings show that, within the observed range of rearing conditions, environmental and feeding covariates carried only marginal information about maturation timing beyond what is captured by elapsed rearing day, and they motivate wider covariate ranges, multistate transition modelling, and additional substrate covariates before an RSF-based decision-support tool for circular organic waste management can be recommended for operational use

Beyond Accuracy: Demographic Feature Sufficiency in Imbalanced Transmission Preference Prediction

Jeffry jeffry — 2026-10-30

Predicting customer preference from limited demographics can appear successful when target classes are highly skewed. This study evaluates whether normalized age and gender indicators are sufficient to predict vehicle transmission preference and whether conventional accuracy masks minority-class failure. A customer dataset of 223,042 records contains MANUAL 178,943, AUTOMATIC 35,810, CVT 8,283, and E-CVT 6. Because E-CVT is too sparse for reliable stratified validation, it is retained for descriptive analysis but excluded from the three-class predictive comparison. We combine association tests, mutual information, an empirical pattern-oracle analysis, and five-fold stratified cross-validation of a majority baseline, multinomial logistic regression, cost-sensitive logistic regression, and class-weighted random forest. Standard logistic regression reproduces the majority baseline, with accuracy 0.8023 but balanced accuracy 0.3333 and MCC 0.0000. Class-weighted random forest lowers accuracy to 0.4946 but improves balanced accuracy to 0.4535, Macro-F1 to 0.3399, and MCC to 0.0969. Age and gender jointly provide only 0.0300 bits of information, approximately 3.5 percent of target entropy. The findings show that high accuracy primarily reflects class prevalence; richer behavioral, product, and transaction features are required for practically useful preference prediction.

Implementation of a Movie Recommendation System Using Collaborative Filtering and the K-Nearest Neighbors Technique

Masnur Masnur — 2026-10-30

The rapid growth of video services has greatly expanded the number of available films, making it increasingly difficult for users to identify titles that match their preferences. This study implements a web-based movie recommendation system using collaborative filtering and the K-nearest neighbors (KNN) technique to generate personalized recommendations from similarities among user preferences. An experimental quantitative design was adopted through requirements analysis, system design, implementation, testing, and evaluation. User ratings were used to calculate inter-user similarity, after which KNN selected the most similar users as the basis for recommendation generation. The system was evaluated using black-box testing and a five-point Likert-scale questionnaire covering system performance, information use, system use, and user satisfaction. All functional test scenarios were classified as valid, while the questionnaire variables received mean scores in the good category. The findings indicate that the system can provide relevant recommendations through an interface that users perceive as effective and easy to use. Collaborative filtering combined with KNN therefore offers a practical approach for helping users find films that correspond to their preferences while improving the service quality of a web-based recommendation platform.

Comparing Four Classifiers for Imbalanced Computer Engineering Student Feedback

Kharis Hudaiby Hanif — 2026-10-30

Open-ended course evaluation comments provide detailed information about students' learning experiences, but large volumes of text are difficult to review consistently by hand. This study compares Decision Tree, Support Vector Machine (SVM), Random Forest, and Gradient Boosting for sentiment classification of Computer Engineering student comments. The initial dataset contained 1,075 comments. Compound-score labeling produced 1,065 neutral comments (99.07%), five negative comments (0.47%), and five positive comments (0.47%). After incomplete records were removed, 415 observations remained for modeling, and the test set contained 207 observations. Before SMOTE, Random Forest and SVM achieved 98.06% accuracy, while Decision Tree and Gradient Boosting achieved 97.58%. After SMOTE, SVM provided the best balance, with 98.55% accuracy, 66.18% macro precision, 66.67% macro recall, and 66.42% macro F1-score. Although overall accuracy was high, the positive class was not detected. Therefore, SVM was the most balanced model in this experiment, while macro-averaged metrics and class-wise errors remain essential when evaluating highly imbalanced sentiment data

Integration Of Fingerprint Attendance Systems And Iot-Based Computer Vision For Employee Presence Monitoring

Aang Kunaefi — 2026-10-30

The fingerprint-based attendance system at the Makassar Public Health Laboratory (BLKM Makassar) has limitations because it only records administrative attendance and cannot verify the physical presence of employees in the workplace. This study aims to design and implement an integrated fingerprint attendance and Internet of Things (IoT)-based computer vision system to support automated employee presence monitoring. The system uses two ESP32-CAM units installed in the lobby area and connected to a Flask Python server and MariaDB database. Image processing employs Histogram of Oriented Gradients (HOG) for face localization, dlib 68 facial landmarks for face alignment, and ResNet-34 to generate 128-dimensional face embeddings. Face identification is performed using Euclidean Distance with a threshold of 0.45 and a Time Window Buffer mechanism during monitoring cycles at 10:00 and 14:00 WITA. The results show that image transmission through HTTP POST operates with HTTP 200 OK and a latency of approximately 80 ms. Under normal walking conditions, the system achieved a 100% detection success rate. From 150 dynamic tests, the multi-camera system achieved an overall detection success rate of 63.33%, while Camera 1 and Camera 2 achieved 60.67% and 40.00%, respectively. The system achieved a 74% success rate under fast walking conditions, which decreased to 16% during running due to motion blur and lighting conditions. The integration of fingerprint attendance and computer vision provides supporting information for identifying discrepancies between administrative attendance records and the visual presence of employees within the monitored camera-covered area.

Stability and Significance Analysis of Daily Food Price Forecasting Models

Gus Nanang Syaifuddiin — 2026-10-30

Food price forecasting has been widely studied using machine learning, yet most published work reports accuracy from a single data split without a baseline comparison or a significance test. This study evaluates the stability and statistical significance of such comparisons. Eleven models comprising four baselines, two statistical models, and five machine learning algorithms were tested on eight strategic food commodities in East Java Province using 2,380 daily observations from April 2017 to September 2026 obtained from the Strategic Food Price Information Centre of Bank Indonesia. Evaluation employed six-fold expanding window validation with a thirty-step horizon, using a scaled error measure and a skill score relative to a naive benchmark as primary metrics, followed by Friedman and Nemenyi testing. No model proved significantly superior to the naive benchmark, which itself ranked third among the eleven models. Rankings proved highly unstable: the standard deviation across folds reached 35 to 106 percent of mean values, the best model changed for half of the commodities after cleaning only 0.1 percent of observations, and percentage error correlated negatively with the skill score. These findings indicate that claims of model superiority require baseline comparison, multi-fold validation, and significance testing.

Javanese Script Classification with YOLO26 and Convolutional Filter Visualization

Harry Mardika — 2026-10-30

Javanese script is a cultural heritage whose use is narrowing, while automatic recognition is held back by visual similarity between characters. This study evaluates the YOLO26 architecture on Javanese script classification, records the change in convolutional kernel weights across the eleven architecture blocks, and tracks prediction probabilities on visually similar characters. Secondary data of 2,659 Nglegena script images from 20 classes came from Kaggle. Each image passed through Otsu's binarization, diacritic-safe bounding, resizing to 224x224 pixels, and letterboxing; the training partition was then expanded through augmentation. The YOLO26 nano variant was trained in PyTorch with the AdamW optimiser and stopped at the tenth epoch through early stopping. Kernel weights were read from every convolutional module each epoch. Testing on 25 images produced 96% accuracy with a macro average F1-score of 0.9733, and 18 of the 20 classes were recognised without error. Blocks 1 to 3 recorded edges, corners and curves at larger weight values, blocks 4 to 9 produced denser representations with weights closer to zero, block 10 concentrated on class-separating areas, and block 11 converted them into class scores. By the tenth epoch most classes stayed above 99%, while 'ha' reached 85%.

Development of an Adaptive Entropy-Based GLCM Feature Extraction Method for Defect Detection in Welding Images

Sherly Agustini — 2026-10-30

Quality inspection of welding joints is fundamental to ensuring structural integrity and reliability, particularly in high-risk industrial sectors. The use of image processing for welding defect detection offers a more objective and consistent approach than visual interpretation; however, the performance of detection systems strongly depends on the ability of feature extraction methods to represent image texture characteristics. Gray Level Co-occurrence Matrix (GLCM) is a widely used texture feature extraction method because it can quantitatively and interpretably represent spatial relationships among gray levels. Nevertheless, static GLCM configurations may limit its representational capability when images exhibit low contrast, non-uniform intensity distributions, or small and complex defect characteristics. This study aims to develop an adaptive entropy-based GLCM feature extraction method through the integration of Adaptive ROI, ROI Entropy, and Feature Fusion. The development adopts a Research and Development (R&D) approach. The GLCM-Entropy method was subsequently tested and compared with GLCM Before Development using 500 radiographic images obtained through X-ray inspection and evaluated using a confusion matrix. The test results show that GLCM Before Development achieved an accuracy of 91.40%, whereas GLCM-Entropy achieved 96.20%, with F1-scores of 87.09% and 97.42%, respectively. These results demonstrate that the integration of global and local texture information in GLCM-Entropy improves classification performance compared with GLCM Before Development.

Early Detection of Potential Tip Burn and Frog Eye in Hydroponic Lettuce Based on Environmental Conditions Using Naive Bayes on ESP32

Fadhlan Muhammad — 2026-10-30

Lettuce (Lactuca sativa L.) is a high-value horticultural commodity susceptible to tip burn, a calcium deficiency disorder triggered by high temperature and low humidity, and frog eye, a fungal leaf spot caused by Cercospora canescens that thrives under low temperature and high humidity. Visual inspection detects both only after damage has occurred. This study developed an Internet of Things early detection system that classifies environmental risk conditions using Gaussian Naive Bayes embedded directly on an ESP32 microcontroller. The system integrates DHT22, pH, and TDS sensors and sends automatic Telegram notifications. Environmental data were collected for 31 days at a hydroponic farm in Makassar, yielding 1,005 valid records after preprocessing. The model was trained in Python with scikit-learn, and its parameters were deployed on the ESP32 as embedded machine learning. The simulation model achieved 95.56 percent accuracy, while the real-data model achieved 94.03 percent accuracy on 201 test samples, with a five-fold cross-validation accuracy of 92.91 percent and a spread of only 0.12 percent across four validation methods. Field visual validation showed 52 percent agreement, indicating that the system is best suited as an early warning of conducive conditions rather than as a symptom diagnosis tool.

Keywords—Embedded Machine Learning, ESP32, Frog Eye, Naive Bayes, Tip Burn

DNN versus CNN for Abusive Audio Classification in YouTube Shorts

Dwi Atmoko — 2026-10-30

Audio-based abusive language in short-form videos is difficult to moderate when the spoken content does not appear in captions. This study compares a deep neural network (DNN) trained on a 183-dimensional engineered feature vector with a convolutional neural network (CNN) trained on 64 by 87 Mel-spectrograms for classifying eight Indonesian abusive keywords. From 923 YouTube Shorts, 272 segments were manually curated, and 151 source clips were augmented into 906 samples standardized to 2 seconds at 22,050 Hz. Evaluation used stratified 5-fold cross-validation, a 70:30 stratified holdout, class-wise metrics, and confusion matrices. The DNN achieved a mean cross-validation accuracy of 87.08 percent with a standard deviation of 2.12 percentage points, and a holdout accuracy of 83.82 percent with a macro F1-score of 83.24 percent. The CNN reached a mean cross-validation accuracy of 81.33 percent with a standard deviation of 30.48 percentage points, but only 13.97 percent holdout accuracy and a 4.00 percent macro F1-score, indicating severe instability and class collapse. The engineered-feature DNN was therefore more stable in this small-data setting, although the results do not establish an intrinsic architectural superiority. Limitations include synthetic augmentation, class imbalance, single-rater curation, an isolated eight-word vocabulary, and no transfer-learning or full-video evaluation.

Coordinate Identity Features for Unsupervised Detection of Attendance Location Fraud

Nur Trismi Auliah — 2026-10-30

Web-based attendance systems cannot read the mock location flag of a user device, thus, the coordinates produced by spoofing applications pass geofence validation unnoticed. This study asks whether detection can be moved from the device to the historical pattern, and which unsupervised detector suits the data that result. Eleven composite spatio-temporal-behavioural features were engineered from the attendance log of a corporate human resource information system in Indonesia, three of which measured how often an employee repeated the same coordinate at six-decimal precision. Isolation Forest, a one-class support vector machine, and a long short-term memory autoencoder were compared under twelve-fold walk-forward cross-validation on 249,818 records from 212 employees, evaluated against pseudo-labels derived from seven documented attendance rules. The stored models were then applied to 18,711 later records, and the leading indicator was confirmed with ten employees through a blinded matched case-control interview. The one-class support vector machine led on every accuracy metric, reaching a precision-recall area of 0.5367 and a Matthews correlation coefficient of 0.2541, and its ranking held across three calibration scenarios. Removing the coordinate repetition features cost 74.2 percent of the discriminative margin, while the field confirmation separated flagged from comparison employees completely.