Jurnal Sistem Cerdas

Analyzing Bias Trade-Offs in Movie Review Sentiment Analysis using a BERT - SVM Enhanced Model

Vany Eka, Hastari Utama — Sun, 19 Apr 2026 00:00:00 +0700

Sentiment analysis of movie reviews often exhibits genre-based bias, where model performance varies significantly across subgroups—an issue that standard accuracy metrics can mask. To address this, we propose a novel fairness-aware hybrid model, BERT-SVM (Fairness-Tuned), which integrates sample re-weighting focused on the lowest-performing genre into the BERT-SVM pipeline. Using a public IMDb movie review dataset from Kaggle, we first train a standard BERT-SVM model and identify Horror as the weakest-performing genre (accuracy: 72.3%, vs. overall 89.6%). We then apply targeted re-weighting to upsample underrepresented or misclassified Horror samples during training. The Fairness-Tuned model reduces the accuracy gap by 62%, raising Horror genre accuracy to 83.1% while maintaining strong overall performance (87.4%). This work not only quantifies the fairness–accuracy trade-off but also demonstrates that lightweight, genre-specific bias mitigation within a hybrid architecture can effectively enhance equity without drastic model redesign—highlighting the value of explicit fairness evaluation in NLP applications

Deep Learning-Based Classification of Fetal Head Abnormalities from Ultrasound Images Using EfficientNet-B3

Galih Hendro Martono, Neny Sulistianingsih — Fri, 26 Dec 2025 00:00:00 +0700

Fetal brain abnormalities represent a critical concern in prenatal diagnostics due to their significant impact on neonatal survival and neurological development. Conventional ultrasound (USG) screening relies heavily on expert interpretation, which can be time-consuming and prone to subjectivity. To overcome this constraint, this research develops an automated classification approach employing deep learning techniques to recognize fetal head abnormalities captured through ultrasound scans. The dataset, obtained from a publicly available Kaggle repository, comprises fourteen diagnostic categories, including Arnold Chiari Malformation, Arachnoid Cyst, Cerebellar Hypoplasia, Holoprosencephaly, and Ventriculomegaly variations, among others. Each ultrasound image was subjected to a series of preprocessing operations, such as resizing to 224×224 pixels, applying normalization, and performing data augmentation, to enrich feature variability and strengthen the model’s generalization capability. A pretrained EfficientNet-B3 architecture was fine-tuned for multi-class classification, with the fully connected layer adapted to predict fourteen distinct abnormality classes. Model training was conducted for ten epochs using the Adam optimizer and cross-entropy loss function, with performance evaluated via training loss and validation accuracy metrics. The results demonstrate rapid convergence, with training loss decreasing from 1.7055 in the first epoch to 0.0387 in the final epoch. Concurrently, validation accuracy improved from 79.60% to a peak of 91.37%, indicating strong generalization capability. The consistent upward trend in accuracy and the downward trend in loss confirm the model’s stability and effective learning behavior. Overall, the proposed EfficientNet-B3–based approach achieves high accuracy and robustness, highlighting its potential as an assistive tool for automated prenatal diagnosis of fetal brain abnormalities

Logistic Regression Model for Predicting SNBP Admission Based on Academic Data

Daffa Naufal Rahimi, Yusuf Ramadhan Nasution — Thu, 16 Apr 2026 00:00:00 +0700

The National Selection Based on Achievement (SNBP) is a crucial pathway for prospective students to access higher education; however, the uncertainty surrounding admission outcomes often causes anxiety among prospective students. This study aims to develop an SNBP admission prediction model based on logistic regression using academic data from students at the Darul Arafah Raya Islamic Boarding School. The method used is binary logistic regression with parameter estimation via the Newton-Raphson method. The research data consists of 261 academic records of students from 2022 to 2025, divided into training and testing datasets. Model evaluation was conducted using accuracy, precision, recall, F1 score, and AUC-ROC metrics. The results show that the model achieved convergence at the seventh iteration with an accuracy rate of 81.25 percent. The precision and recall values were 82.35 percent, respectively, while the AUC-ROC value was 0.9049, which falls into the “good classification” category. It can be concluded that the logistic regression model is effective for predicting SNBP graduation based on average report card scores and is suitable for implementation as a decision support system for students in estimating their admission chances.

Optimization of ESP32-Based Living Room Security System and PIR Sensors, with FTDI through real-time notifications

ahmad Supyan, Nur Azizah, Firman Jaya — Sat, 14 Feb 2026 00:00:00 +0700

This study addresses the problem of undetected falls and hazardous incidents in household living rooms, especially for child and elderly users who are prone to slipping on wet and narrow surfaces. The study aimed to design and implement a smart living room safety monitoring system using ESP32 microcontrollers, PIR sensors, magnetic sensors, load cells with HX711, and MPU6050 connected to the Internet of Things to provide real-time notifications to caregivers via mobile apps. This methodology follows a prototype-based IoT engineering approach, starting with a literature review and needs analysis, followed by hardware-software design, prototyping, iterative testing, and final evaluation in a simulated living room environment for various fall scenarios. The experimental data consisted of PIR logs, weight changes, system response time, and environmental conditions, which were statistically analyzed to determine the system's accuracy, reliability, and responsiveness. The results showed that the prototype was able to detect suspicious movement patterns and falls with good accuracy and trigger local alarms and Telegram notifications within about 2–3 seconds, while still functioningin poor living room conditions. It can be concluded that the proposed system meets the research objectives ofa low-cost and privacy-preserving living room safety solution for smart homes, with future work directed at integrating machine learning-based fall detection and expanding communication options beyond WiFi to improve resilience in various residential environments.

An Intelligent IoT-Based Water Quality Monitoring and STORET Index Prediction System Using Random Forest

Yohanes Nugroho, Dewi Indriati Hadi Putri, Mahmudah Salwa Gianti — Thu, 16 Apr 2026 00:00:00 +0700

Hygiene sanitation water quality fluctuates due to environmental dynamics, yet conventional monitoring systems generally lack the predictive capabilities compliance with health standards (Permenkes No. 2 of 2023). This study aims to develop an intelligent Water Quality Monitoring System (WQMS) capable of predicting water quality status based on the STORET index using the Random Forest algorithm. [Methods] The proposed system integrates an ESP32-S3 microcontroller with calibrated low-cost sensors for real-time data acquisition. To ensure data integrity, regression was applied for sensor calibration, while the STORET method was utilized to determine pollution levels and water feasibility. A Random Forest regression model was then trained using these processed datasets to classify water quality status. Experimental results demonstrated high hardware precision, achieving Mean Absolute Percentage Error (MAPE) values of 5.38% for pH, 2.24% for TDS, and 0.22% for the Flow Meter. Furthermore, the Random Forest model exhibited superior predictive performance, yielding a Coefficient of Determination () of 0.9977, a Mean Absolute Error (MAE) of 0.2213, and a Root Mean Squared Error (RMSE) of 0.5144. These findings indicate that the integrated system effectively combines accurate sensing with robust predictive modeling. Consequently, the system is categorized as highly capable of providing real-time insights and early warnings, offering a significant improvement over traditional monitoring methods for public health safety.

Real-Time Telemetry Based Monitoring System for Energy Efficiency Evaluation of Scheduled Dimming in Office Corridor Lighting

Muhammad Risqi Nuryana, Veronica Windha Mahyastuty, Ridwan Satrio Hadikusuma — Sat, 18 Apr 2026 00:00:00 +0700

In many office buildings, corridor lighting systems are commonly operated at full brightness continuously, regardless of occupancy conditions, resulting in unnecessary energy consumption. This study proposes and evaluates a real-time telemetry-based monitoring system to assess the energy efficiency of a scheduled dimming strategy for office corridor lighting. The developed system integrates dimmable LED luminaires with a telemetry unit capable of transmitting real-time illuminance and energy consumption data for monitoring and analysis. A time-based dimming schedule of 25%, 50%, and 75% output levels was implemented in an office corridor environment. Illuminance measurements were collected at five different points along the corridor, while electrical energy consumption was recorded continuously over a seven-day observation period through the telemetry monitoring platform. The results indicate that even at the lowest dimming level (25%), the corridor maintained an average illuminance of 100 lux, which remained within acceptable lighting standards for pedestrian circulation. Telemetry data further demonstrated that the scheduled dimming strategy reduced weekly energy consumption by approximately 64% compared to continuous full operation (0.711 kWh reduced to 0.253 kWh). These findings confirm that real-time telemetry monitoring enables accurate performance evaluation of lighting control strategies while ensuring compliance with visual comfort requirements. The study highlights the potential of telemetry-based lighting monitoring systems as an effective approach to optimize energy use, minimize over-lighting, and support data-driven energy management in office buildings.

ECRM Strategy In Improving Services And Sales Of Electronic Products

Qevin Dwi Rafitrah, Dewi Anggraeni, Ruri Ashari Dalimunthe — Thu, 02 Apr 2026 00:00:00 +0700

Business competition in the era of globalization demands companies to continuously adapt, one of which is through the implementation of Electronic Customer Relationship Management (E-CRM). Mega Cell Sei Piring Pulau Rakyat, a retail store for electronic products, faces challenges in the form of inconsistent sales fluctuations and customer management that is still manual. If this unstable sales pattern continues without improvement, it has the potential to disrupt profitability and create a risk of loss. This study aims to implement an E-CRM strategy as a strategic solution to improve customer service, optimize sales management, and maintain the company's revenue stability. This qualitative research produces a design and prototype of a web-based E-CRM system (PHP and MySQL) that enables transaction recording, product management, promotion management, and integrated sales reporting. The E-CRM system that has been built facilitates market segmentation and better customer interaction, thus it is expected to improve operational efficiency, reach a wider market through online information access, and encourage customer attraction and loyalty through data-based promotional offers

A Machine Vision–Based Automated Wheel Leak Detection System Using Real-Time Object Detection in the Water Leak Testing Process

Mon, 20 Apr 2026 00:00:00 +0700

Water leak testing in automotive wheel manufacturing has traditionally relied on manual visual inspection of bubble formation, introducing subjectivity and limiting repeatability in quality assurance processes. This study developed and experimentally validated a real-time leak detection system based on machine vision, directly integrated with an industrial water leak tester platform. A dataset comprising 686 annotated images was constructed from recorded operational testing sequences and partitioned into 80% training and 20% validation subsets. The network was trained for 150 epochs and deployed within an integrated framework incorporating temporal decision logic and automated event logging to ensure deterministic classification under continuous video streaming. Experimental validation was conducted across five scenarios (A–E), including high-leak, low-leak, no-leak, and in-situ operational testing conditions, totaling 100 trials. The aggregated confusion matrix yielded 60 true positives and 40 true negatives with zero false positives and false negatives, resulting in accuracy, sensitivity, specificity, precision, and F1-score values of 1.0 within the evaluated domain. Receiver operating characteristic and precision–recall analyses confirmed strong class separability and stable decision boundaries. Although the results demonstrated high discriminative performance under controlled and operational settings, further large-scale validation under heterogeneous industrial environments is required to fully assess long-term robustness. The proposed framework provided an automated, objective, and real-time inspection solution aligned with Industry 4.0 principles for intelligent manufacturing systems.

Analysis of Diabetes Classification Performance Improvement Using Ensemble Bagging and K-Fold

Mawardi Kudin, Abd Salam At Taqwa, Angga Kurniawan, Chairi Nur Insani — Fri, 17 Apr 2026 00:00:00 +0700

Diabetes mellitus represents a long-term metabolic disorder whose global incidence continues to rise, making precise early identification essential to minimize severe complications. Machine learning techniques have been extensively utilized for diabetes classification; however, single-model approaches often suffer from performance constraints, such as susceptibility to overfitting and high variability in prediction outcomes. To address these challenges, this research introduces a bagging-based ensemble learning strategy integrated with K-Fold Cross Validation to enhance both predictive accuracy and model robustness. The study employs the Pima Indians Diabetes Dataset, which contains 768 patient records described by eight clinical features and one outcome variable. Eight classification methods—Logistic Regression, K-Nearest Neighbors, Support Vector Machine, Decision Tree, Random Forest, Naïve Bayes, Gradient Boosting, and XGBoost—were assessed individually and within the proposed ensemble framework. Model effectiveness was measured using accuracy, precision, recall, and F1-score derived from the confusion matrix. The findings indicate that the ensemble bagging approach generally strengthens model stability and yields improvements in accuracy and precision across most algorithms. Notably, K-Nearest Neighbors and XGBoost demonstrated the most stable gains following ensemble integration. Nevertheless, enhancements in precision were frequently associated with a reduction in recall, reflecting a trade-off in identifying positive cases. In summary, the integration of bagging and K-Fold Cross Validation provides a more resilient and dependable classification model, offering strong potential for supporting clinical decision-making in early diabetes detection.

Implementation of Content-Based Cosine Similarity Algorithm with TF-IDF and SBERT for Movie Recommendation

Eliata Zefanya Irabela, Norhikmah — Sun, 26 Apr 2026 00:00:00 +0700

The number of films continues to increase on streaming platforms often makes users confused in deciding which film to watch. To overcome this research develops content-based movie recommendation system. Representation of the film information obtained through the application of TF-IDF and SBERT to genre and synopsis data. Cosine similarity is used to calculate the closeness between representations. The performance system is then evaluated through the Precision@K, MAP@K, and Recall@K metrics. From the test results, hybrid approach shows better performance more stable than single method. With a MAP value reaching 0.95 Recall 0.95 dan Precission 0.71 . In the future, the development system will still possible by utilizing other types of data, including user interaction data.

Design and Development of a Web-Based Laundry Management System (Case Study of Yamus Laundry)

Ayu Latifah, Moch Idham Hanafi — Tue, 28 Apr 2026 23:34:28 +0700

Yamus Laundry faces significant operational challenges due to its manual process for recording customer, transaction, and inventory data, which is prone to errors, data loss, and service disruptions. The objective of this research is to design and build a web-based laundry management system to serve as a centralized platform to address these problems. The system was developed using the Rational Unified Process (RUP) method with system modeling based on the Unified Modeling Language (UML), and it was implemented using the PHP programming language and the Laravel framework. This research resulted in a functional web application that has successfully passed Black Box testing. The system includes key features such as integrated customer data management, transactions with automatic quota deduction, and an inventory module with low-stock notifications. The implementation of this system has been proven to reduce the risk of recording errors, lighten the staff's workload, and improve data accuracy, allowing all operational activities at Yamus Laundry to run in a more organized and efficient manner.