Formosa Journal of Computer and Information Science

IoT-Based Multi-Sensor Fusion for Goat Behavioral Pattern Recognition Using K-Means Clustering in a Smart Farming Environment

Yudhistira Pratama — 2026-03-30

Monitoring goat behavior in commercial farms typically relies on direct observation, which does not scale and misses conditions that develop gradually. This study deployed an eight-sensor IoT network across two zones of a slatted-floor goat pen in North Sumatra, Indonesia, and applied K-Means clustering to 49 days of sensor data. After a systematic data cleaning step that removed sensor dropouts, ADC saturation events, and an isolated methane spike, 213,704 records were retained (98.6% of raw data). K-Means with K=8 on the cleaned dataset yielded a Silhouette Score of 0.297 and Davies-Bouldin Index of 1.177, identifying eight behavioral and environmental states without a dedicated anomaly cluster. Results include two heat stress levels (THI means 90.7 and 92.1), three nocturnal resting states differentiated by waste pit gas concentration, a daytime active-vocal state, and an evening post-feeding fermentation peak.

Comparative Study of Machine Learning Models for Sentiment Analysis of Amazon Product Reviews

Tri Noviantoro — 2026-03-30

This research presents a comparative analysis of four popular sentiment classification models: Naive Bayes, Support Vector Machine (SVM), Long Short-Term Memory (LSTM) networks, and Bidirectional Encoder Representations from Transformers (BERT). The models are evaluated using the Amazon Product Reviews dataset based on their ability to classify sentiments into positive or negative categories. The results show that BERT outperforms the other models in accuracy, precision, recall, and F1-score, demonstrating its superior ability to capture complex contextual relationships in text. LSTM performed well, particularly in recalling positive sentiments, but was outperformed by BERT overall. Conversely, Naive Bayes and SVM exhibited lower accuracy and higher false positive rates, highlighting their limitations in handling nuanced, context-dependent text. This study emphasizes the trade-offs between traditional machine learning models and advanced deep learning techniques.

Building an Annotated Corpus of Advice-Giving in Indonesian Thesis Supervision for Educational Text Mining

Elok Putri Nimasari — 2026-03-30

While educational text mining has widely examined student feedback and institutional evaluation, little attention has been paid to advice-giving in thesis supervision as an interactional and power-relational practice. Therefore, this present study aims to analyze and build a domain-sensitive annotated corpus of advice-giving in Indonesian thesis supervision for future educational text mining. Using a qualitative-informed corpus development research design, the study collected and analyzed 155 annotated utterances drawn from authentic thesis supervision transcripts across Indonesian universities. The results identified six advice-giving labels classified into three interactional modes: power-over, power-gaining, and power-maintaining following Zhang and Hyland’s theoretical of power and roles. Cohen’s Kappa reached 1.00, indicating perfect annotation agreement. The corpus contributes a reliable methodological foundation for AI-assisted analysis of supervisory discourse and inclusive academic supervisory.

Interpretive Practices of PLC Based Automation in Industrial Production Systems under Dynamic Operational Conditions

Ferdianto Tangdililing — 2026-03-30

This study examines interpretive practices in PLC-based automation, particularly how real-time data supports adaptive decision-making in PLC–SCADA manufacturing systems. Using a mixed-method approach, the findings reveal that system effectiveness depends not only on deterministic control logic but also on the integration of sensor data, HMI visualization, and adaptive control. Under dynamic conditions, effective data interpretation enhances production efficiency, reduces downtime, and accelerates response to disruptions. Furthermore, the integration of PLCs with IoT and data analytics improves system flexibility and reduces decision ambiguity. The study concludes that successful PLC automation relies on the synergy between control technology, data interpretation, and human–machine interaction, contributing to the development of more adaptive and intelligent production systems.

Transient Analysis of EVA Foam Damping and Spatial Optimization of FSR Sensors on Shain Guards

Fahrizal Akbar Herbhakti — 2026-03-30

Embedding Force Sensitive Resistors (FSR) into EVA foam for smart shin guards is hindered by stiffness mismatch between the soft matrix and rigid sensor. This study determines the optimal embedment depth that balances signal fidelity and structural integrity. Coupled multiphysics FEM simulations (COMSOL) employing a hyper elastic Mooney-Rivlin model and piezoresistive equations were run under a 1500 N peak Gaussian impact pulse. At 2 mm depth, sensitivity reached –85% ΔR but shear stress peaked at a critical 42.5 MPa; 8 mm depth was very safe (12.4 MPa) but gave a weak –25% ΔR. The optimum depth was 5 mm, yielding 24.8 MPa shear stress, –62% sensitivity, 2.5 ms latency, and high SNR. Sensitivity analysis and numerical optimization confirmed this sweet-spot. The computational framework provides precise parameters for manufacturing IoT-enabled smart shin guards.

Effects of Scratch Gamification with MDA on Students’ Engagement and Learning Outcomes

Agustinus Sembiring — 2026-03-30

This study aims to examine the effect of Scratch-based gamification using the MDA (Mechanics–Dynamics–Aesthetics) model on students’ engagement and learning outcomes in primary education. A quasi-experimental method with a pretest–posttest design was applied to 115 students. Data were collected through tests and questionnaires and analyzed using paired sample t-tests and descriptive analysis. The results showed a significant improvement in learning outcomes, with a mean pretest score of 56.84 and posttest score of 71.03 (t = -26.57; p < 0.001). In addition, students’ engagement was categorized as high (mean = 3.73). These findings indicate that Scratch-based gamification integrated with the MDA model is effective in improving learning quality.

Implementation of Decision Support System with Simple Additive Weighting (SAW) Method for Determination of Social Assistance Recipients: A Case Study in Ciledug, Tangerang

Tengku Rafi Syahrial — 2026-03-30

This can lead to dissatisfaction among the community and reduce the effectiveness of social assistance programs. This research aims to develop a Decision Support System (SPK) based on the Simple Additive Weighting (SAW) method to ensure a fairer, more objective, and efficient distribution of social assistance. The SAW method is used to assess aid recipients based on five main criteria, namely family income, number of dependents, home conditions, employment status, and age of the head of the family. Each criterion is weighted according to its level of importance, and the data obtained from the respondents is processed through a process of normalization and final score calculation to determine eligible beneficiaries. The implementation of this method in RT 02, RW 06, Ciledug, Tangerang City, showed a significant increase in the efficiency and accuracy of the selection of social assistance recipients.

The Application of Naive Bayes in Analyzing Public Sentiment Toward the Performance of the North Sumatra Regional Government in Handling Flash Floods

Rivaldo Siburian — 2026-03-30

This study analyzes public sentiment towards the performance of the North Sumatra Regional Government in handling flash floods using the Multinomial Naive Bayes algorithm. A total of 1,132 opinion data points were collected from social media and news portals through web crawling from November 2025 to February 2026. Sentiment labeling was performed using a lexicon-based approach with the InSet dictionary. Classification results showed a dominance of negative sentiment at 88.4%, focusing on slow emergency response. Model evaluation with an 80:20 data split yielded 89.43% accuracy and an F1-Score of 0.844 for Naive Bayes, while SVM achieved the highest F1-Score (0.855). This study concludes that AI-based sentiment analysis can serve as an objective instrument for government performance auditing.

Enhancing Business Model Validation Using Artificial Intelligence: Insights from Student Business Model Canvas Analysis

Dimas Setiawan — 2026-03-30

This study aims to identify common weaknesses in students’ Business Model Canvas (BMC) and examine the role of Artificial Intelligence (AI) in improving early-stage business validation. A qualitative descriptive approach was employed using aggregated and anonymized data from 30 student business models. Data were analyzed using thematic coding to identify recurring patterns. The findings reveal that 70% of students struggle with unclear value propositions, 63% define overly broad customer segments, and 57% lack structured revenue models. AI-assisted analysis improves clarity, focus, and logical consistency of business models. This study proposes an AI-BMC conceptual framework as a decision-support approach for entrepreneurship learning. The findings contribute to bridging intuitive business ideation with AI-assisted validation.

Comparative Analysis of Traditional Machine Learning and Sequential Deep Learning Models for Spam Email Classification

Harliana Harliana — 2026-03-30

This study compares the performance of traditional machine learning methods and sequential deep learning models for text-based spam classification. The primary issue addressed is the lack of consistent, fair evaluation across these approaches due to variations in datasets, preprocessing techniques, and experimental settings across previous studies. To overcome this limitation, this research proposes a controlled comparative evaluation framework by employing a unified dataset, standardized preprocessing procedures, consistent data splitting, and identical evaluation metrics. The dataset used consists of 5,572 messages with an imbalanced class distribution; therefore, oversampling was applied to the training data to mitigate bias. The evaluated models include TF-IDF-based Logistic Regression as the baseline, as well as Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), and Gated Recurrent Units (GRUs) as deep learning models.

Microservices-Based Open-Source Video Conference Deployment for Optimized Online Learning Infrastructure

Davy Putra Ananda — 2026-03-30

The rapid advancement of information technology has fundamentally shifted the interaction paradigm in education from conventional methods to hybrid learning models. In this context, the availability of stable, real-time communication platforms has become crucial for maintaining the effectiveness of knowledge transfer. This study evaluates the implementation of Apache OpenMeetings v9.0.0 using Docker and WSL2 to provide efficient video conferencing. Using an experimental methodology, system performance was monitored during active sessions. Results show high resource efficiency with a stable CPU utilization of 4.94% and memory usage of 1.339 GiB. The system achieved a rapid startup velocity of 11.1 seconds, proving that containerization offers optimal isolation with minimal overhead. The study concludes that this architecture provides a lightweight, portable, and cost-effective solution for independent communication infrastructure in educational institutions.

Design and Build a Real-Time Based 3 Kg LPG Distribution Information System Using the Laravel Framework at PT. Tasya Gasindo

Yasmin Neylanda Pulungan — 2026-03-30

PT. Tasya Gasindo, an LPG 3 kg distribution agent in Medan, Indonesia, faced significant operational inefficiencies due to the absence of a real-time stock and demand monitoring system. Trucks frequently returned to the depot with unsold cylinders, causing queuing delays and Pertamina fines for unmet daily Delivery Order targets. This study designed and implemented a web-based distribution information system using the Laravel frame-work (MVC architecture, MySQL) providing real-time stock and demand data across 69 distribution bases. The descriptive-qualitative and software engineering approach included observation, interviews, and documentation. Black Box Testing covered 16 scenarios (all passed) with direct user evaluation. Average monthly residual cylinders decreased from 744 units (2022–2024) to 220 units in March–April 2025 a 70.4% reduction. The Fulfill-ment Rate improved from 99.20% to 99.74%, reducing estimated monthly Pertamina fines from Rp22.32 million to Rp6.60 million

Analysis of PT PLN (Persero)'s New Installation Waiting List Using the K-Means Clustering Algorithm

Ernawati Ernawati — 2026-05-21

This study examines the application of the K-means clustering algorithm to analyze new installation waiting list data obtained from the last three months of 2024. Only entries categorized under new installation requests were selected as the primary dataset. The analysis began by determining the optimal number of clusters: a high volume of new installation waiting lists (C1), a medium volume (C2), and a low volume (C3). Data mining processes were carried out using the RapidMiner tool, producing the following results: 6 UIDs/UIWs were classified into the high cluster (C1), 7 into the medium cluster (C2), and 9 into the low cluster (C3). The clustering performance was subsequently validated using the Davies–Bouldin Index, yielding a final score of 0.486, consistent with the RapidMiner output.