CogITo Smart Journal

A Machine Learning-Based Ambiguous Alphabet Recognition for Indonesian Sign Language System (SIBI)

Yoan Purbolingga — 2025-06-30

One of the communication problems in deaf people is the inhibition of verbal communication. This is due to the limited hearing function which has an impact on the imperfection of language sound reception. To communicate with deaf people, extraordinary communication is needed so that the meaning of the conversation can be conveyed properly. Sign language is the main communication medium for deaf people. However, in the use of sign language, there are ambiguous letters, namely “D “,“E“,“M“,“N“,“R“, “S“, and “U“. This research uses the chain code method to identify and reconstruct the shape of hand gesture objects. Then, to solve the problem of ambiguity of alphabet letters, an artificial intelligence method, namely K-Nearest Neighbors (K-NN), is used. The sample used consists of 350 real-time images with variations in object recognition accuracy. Based on the research using chain code and K-NN classification method, it can be concluded that the recognition of ambiguous letters in sign language has 245 training data for K-NN which has 88.76% accuracy, and 105 test data with 90% accuracy. This test data is divided into seven letters: “D“, “E”, “M”, “R” and “U” at 100%, and “N” and “S” at 98.88%.

Prototype of IoT-Based Temperature and Humidity Monitoring and Controlling System for Broiler Chicken Coops

Oktoverano Lengkong — 2025-06-30

Temperature and humidity are critical factors that influence the health and productivity of broiler chickens. Manual monitoring and control of coop conditions are often ineffective and inefficient, leading to a decline in production quality. This research aims to develop a prototype IoT-based monitoring and controlling system for temperature and humidity in broiler chicken coops. The system employs a DHT22 sensor to measure temperature and humidity, a Wemos D1 R1 microcontroller for data processing, and the Blynk application as a user interface for real-time monitoring and notifications. The Evolutionary Prototyping method is applied in the development of this system to allow gradual adjustments based on user needs. Testing results show that the prototype can monitor temperature and humidity in real-time and automatically activate fans or lights when the temperature is outside the optimal range. With this system, farmers can monitor coop conditions remotely, simplifying farm management.

Stress Level Detection System Based on Internet of Things

Ragiel Hadi Prayitno — 2025-06-30

Stress is a critical contemporary health issue, causing significant mental and physical damage. Regular monitoring of a person's stress level is vital for early diagnosis of abnormalities that can lead to chronic diseases in the future. While various stress detection methods exist, this research introduces a novel and accessible IoT-based system that integrates a unique combination of physiological parameters: heart rate, body temperature, and, distinctively, exhaled CO₂ concentration. This approach utilizes low-cost, readily available components, including an Arduino Mega microcontroller, an ESP8266 Wi-Fi module, a Pulse Heart Rate sensor, a DS18B20 temperature sensor, and an MQ-2 sensor to measure respiratory CO₂. The significance of this work is demonstrated through its successful implementation and testing on five young adult subjects. The results establish a clear correlation between the measured biometrics and four distinct stress classifications (severe, moderate, mild, and normal). All data is displayed in real-time on a local LCD and transmitted to the Thingspeak IoT platform for continuous analysis. This study confirms the feasibility of using this novel sensor combination to create an affordable, real-time stress monitoring tool, offering a significant contribution to preventive healthcare through early detection and intervention.

Optimizing Text Classification Using Techniques AdaBoost Ensemble with Decision Tree Algorithm

Marnis Nasution — 2025-06-30

This study presents an optimized text classification framework combining AdaBoost ensemble techniques with Decision Tree algorithms (ID3, C4.5, CART) to address critical challenges in small dataset scenarios (n=795 Indonesian-language reviews). Employing rigorous five-fold stratified cross-validation (random seed=42), we implemented a comprehensive preprocessing pipeline including case normalization, language-specific stemming, and TF-IDF feature extraction. The ensemble model utilized 50 AdaBoost iterations with a learning rate of 1.0, evaluated through multiple performance metrics while accounting for class imbalance effects. Key results demonstrate significant performance enhancements, with the C4.5+AdaBoost configuration achieving 96.72% accuracy (±0.88), representing a 10.6 percentage point improvement over the base C4.5 algorithm. The ensemble approach particularly improved minority class identification, boosting positive sentiment classification F1-scores by 0.28 points while maintaining exceptional neutral sentiment detection (F1-score 0.99±0.00). Comparative analysis revealed consistent advantages across all Decision Tree variants, with accuracy improvements of 18.6% for ID3, 10.6% for C4.5, and 14.2% for CART, alongside reduced performance variance (62-73% decrease). While these findings validate AdaBoost's effectiveness for enhancing Decision Tree stability in small-scale text classification, the study acknowledges limitations regarding sample size constraints and language specificity. The research contributes practical methodologies for sentiment analysis applications while emphasizing the need for validation on larger, more diverse datasets. Future work should explore comparative benchmarking against transformer architectures. Advanced feature representation techniques and multilingual generalization testing. This work provides a reproducible framework for developing robust, ensemble-based text classification systems in resource-constrained scenarios.

Cosine Similarity-Based Evidences Selection for Fact Verification Using SBERT on the FEVER Dataset

Harya Gusdevi — 2025-06-30

The spread of misinformation on digital platforms has emphasized the urgent need for automated fact verification systems. However, selecting the most semantically relevant evidence to support or refute a claim remains a challenge, especially within the widely used FEVER dataset. Traditional approaches like TF-IDF often fall short in capturing the contextual meaning between claims and evidence. This study addresses the problem by comparing TF-IDF with Sentence-BERT (SBERT) in measuring semantic similarity. The novelty of this research lies in embedding both claims and evidence using SBERT, then calculating cosine similarity to quantify their semantic relevance. Before embedding, standard preprocessing steps were applied, including tokenization, stemming, lowercasing, and stopword removal. A quantitative approach is used to compute cosine similarity between claim-evidence pairs using both TF-IDF and SBERT embeddings. Similarity analysis, distribution statistics, and t-tests are conducted to evaluate the methods. The results show that SBERT achieves higher similarity with the “SUPPORTS” category (0.65) and stronger negative similarity with “NOT ENOUGH INFO” (-0.90), compared to TF-IDF (0.49 and -0.62, respectively). SBERT also demonstrates more stable score distributions and significantly higher t-test values across all label comparisons, indicating stronger semantic discrimination. These findings confirm that SBERT outperforms TF-IDF in identifying the most relevant evidence. The new dataset generated can serve as a foundation for future fact verification model development.

Advancing Parking Systems: A Performance Comparison of MobileNet and Canny in License Plate Detection

Suryani Suryani — 2025-06-30

Rapid advancements in technology, particularly in computer science, have driven progress in image processing, which plays a crucial role in daily life. This research focuses on object recognition through vehicle license plate detection, utilizing an image database to address human errors in recording vehicle numbers that can slow down parking system services. An automated system is proposed to enhance parking management, although challenges in accurately segmenting plates remain. Two segmentation methods are compared: the MobileNet architecture and the Canny algorithm. This study aims to evaluate the segmentation accuracy between the two methods. Canny for its edge detection capabilities that reduce noise, and MobileNet for its effectiveness as a deep learning-based approach. The system is implemented using Python, JavaScript, HTML, and CSS to modernize vehicle license plate segmentation. The results show that MobileNet significantly outperforms the Canny algorithm, achieving a lower Character Error Rate (CER) of 18.8%, compared to Canny's 50.96%, across 13 tested license plate samples. This finding demonstrates that MobileNet offers a more reliable and accurate approach for segmenting vehicle license plates, thereby contributing to the development of a more efficient and automated license plate recognition system.

Optimizing Heart Disease Prediction with Random Forest and Ensemble Methods

Imam Al Amin — 2025-06-30

This study evaluates ensemble learning techniques for optimizing heart disease prediction, with a focus on Random Forest due to its robustness in handling complex medical data. The dataset used, "Heart Disease Prediction Dataset" from Kaggle, consists of 270 instances and 13 features like age, cholesterol, and family history. Data preprocessing involved mean imputation for missing values and min-max normalization. The study compares Random Forest with other ensemble classifiers—AdaBoost, Gradient Boosting, and XGBoost—using 10-fold cross-validation and evaluation metrics such as accuracy, precision, recall, and F1 score. Results show that Random Forest outperforms the other models with an accuracy of 87.04%, precision of 85.00%, recall of 80.95%, and F1 score of 82.93%. These findings emphasize Random Forest's ability to maintain prediction stability across various medical attributes and imbalanced data. Although the study highlights Random Forest as a promising method for early heart disease risk prediction, it remains a computational evaluation and requires clinical validation. The results aim to inform the development of predictive tools for enhancing early diagnosis and preventive strategies in healthcare systems.

Evaluation of Various Error Metrics in K-NN Image Classification: A Case Study on Egg Image Processing

Frainskoy Rio Naibaho — 2025-06-30

Image processing is a technique to create an image that appears and can be converted into light that describes 2 dimensions. The K-Nearest Neighbor (K-NN) algorithm is known for its efficiency and effectiveness in classification. Although K-NN is often used, a deep understanding of Mean Squared Error (MSE), Mean Absolute Percentage Error (MAPE), and Mean Absolute Deviation (MAD) especially in the context of image classification, still requires further exploration. This study aims to emphasize and measure the behavior of MSE, MAPE, and MAD when applied to K-NN classification results. In this study, calculating statistical features such as mean, standard deviation, skewness, and kurtosis are extracted from the histogram of chicken egg images. These features are then normalized and used as input for the K-NN algorithm. Classification performance is evaluated using MSE, MAPE, and MAD. The results show that K-NN is able to classify chicken egg images with performance measured by MSE of 0.35114462, MAPE of 0.14237803%, and MAD of 0.52930142. The differences in these values, especially the much smaller MAPE value and in percentage units, underscore the importance of selecting the right metrics according to the needs of interpretation and practical applications. This study provides a clearer understanding of how different error metrics characterize the performance of K-NN in image classification, while also highlighting the need for comparative metric considerations in future research.

Gen Z Business Students’ Intention to Adopt Neobanks: An Extended UTAUT Approach

Andrew Christian Aseng — 2025-06-30

As neobanks emerged as one of exciting services in banking industry in recent years and trying to penetrate young customers, this study attempts to examine whether variables in UTAUT model (performance expectancy, effort expectancy, and social influence) and additional variables such as features, curiosity, and rewards affect intention to adopt neobanks with perceived credibility as mediating variable. A quantitative survey was conducted on 359 Gen Z business students at a private university in North Sulawesi. The data collected were then analyzed using a structural equation model (SEM). The results show that only curiosity and perceived credibility have a positive effect on intention to adopt, while performance expectancy and social influence impact on perceived credibility. In addition, perceived credibility mediates the effects of performance expectancy and social influence on intention to adopt. Some recommendations are given in order to increase the Gen Z intention in adopting neobanks while at the same time improving the credibility of neobanks’ applications.