Sinkron : jurnal dan penelitian teknik informatika
Not a member yet
1261 research outputs found
Sort by
Addressing Class Imbalance in Stunting Classification Using SMOTE Enhanced Random Forest
No full textStunting is a chronic nutritional problem that poses serious long-term effects on children’s health, including impaired physical growth, delayed cognitive development, and reduced productivity in adulthood. Early and accurate detection of stunting is therefore essential to support effective public health interventions and targeted policy implementation. However, one of the central challenges in developing machine learning models for this purpose is the presence of class imbalance in health-related datasets. Such imbalance frequently leads to biased classifiers that perform well on majority classes but fail to identify minority categories, reducing the overall reliability of the system. To overcome this issue, the present study utilized the Synthetic Minority Oversampling Technique (SMOTE) to balance the distribution of classes in a dataset containing 110,000 records. A Random Forest algorithm was then employed as the base classifier, with hyperparameter optimization carried out using the Optuna framework to ensure robustness and generalizability. The experimental results demonstrate that the combined application of SMOTE and Optuna significantly improved classification performance, producing the highest Macro Area Under the Curve (AUC) of 0.9972. This outstanding score indicates the model’s superior ability to distinguish nutritional status categories across both majority and minority classes. The study concludes that addressing data imbalance through oversampling is a fundamental methodological step in constructing fair and effective machine learning systems for stunting detection, ultimately contributing to improved health outcomes and evidence-based policy design
Comparison of ARIMA and GRU Methods in Predicting Cryptocurrency Price Movements
No full textThis study compares the effectiveness of the ARIMA and GRU models in predicting Bitcoin price movements, addressing the need for reliable predictive tools amidst the high volatility of the cryptocurrency market. Previous research has highlighted the strengths of each model in financial forecasting: ARIMA for short-term, stationary data and GRU for capturing complex temporal patterns. The purpose of this study is to evaluate which model performs better in the context of Bitcoin price prediction, offering insights for investors to minimize risks and enhance decision-making in this unpredictable market. The research methodology involves applying both models to Bitcoin price data and comparing their accuracy using the Mean Absolute Percentage Error (MAPE) across various forecasting intervals. Results indicate that GRU achieves higher accuracy in long-term forecasts, while ARIMA performs optimally for shorter time frames. However, both models demonstrate limitations, especially as the prediction horizon extends, underscoring the inherent challenges of cryptocurrency price forecasting. These findings suggest that GRU may be better suited for longer investment horizons, while ARIMA remains effective for short-term predictions. The conclusions affirm the potential of using these models selectively to align with specific investment strategies in cryptocurrency markets, although further research is recommended to improve predictive accuracy under evolving market conditions
Stock Price Prediction Using TCN-GAN Hybrid Model
No full textThe stock market plays a vital role in national economies, offering significant profit opportunities for investors while exposing them to substantial risks due to market uncertainties. Stock prices often experience significant fluctuations, making accurate prediction a challenging task. Temporal Convolutional Network (TCN) and Generative Adversarial Network (GAN) are the deep learning method proposed for this research. The purpose of this research is to analyze how well the TCN-GAN model predicts stock prices. Previous researches show both TCN and GAN perform well on time series data. TCN excels in analyzing time-series data while GAN enhances training by generating realistic simulations. By combining the strength of both models, this approach aims to enhance stock price prediction accuracy. The proposed model uses TCN as the generator within the GAN framework and a Multilayer Perceptron (MLP) as the discriminator. TCN handles the prediction task and is trained using the GAN model. The model is trained over 500 epochs, with a learning rate of 0.0004 for the generator and 0.0001 for the discriminator. During each epoch, the generator is updated twice to enhance its performance. The resulting model achieves a MAPE score of 2.16% and an RMSE score of 814.25 on the testing dataset, demonstrating excellent performance in stock price prediction despite significant price variations
Thyroid Disease Prediction Using Random Forest with KNNImputer for Missing Values
No full textThyroid disease is a health dysfunction that requires immediate and accurate diagnosis. This research seeks to design a classification model based on the Random Forest algorithm to detect the type of thyroid disease utilizing data from the UCI Repository. In the data processing stage, KNNImputer is used to handle missing data by calculating the average value of the nearest neighbors based on Euclidean distance, thus ensuring better data quality for model training. The developed model was evaluated utilizing the confusion matrix, which showed an accuracy of 98%, with precision, recall, and F1 score values reached 98% based on weighted avg.These results corroborate that the proposed model is highly reliable in detecting various types of thyroid diseases, such as Negative, Hypothyroid, and Hyperthyroid. This research makes an important contribution to the application of data mining technology for medical diagnosis, while proving that optimal data processing through methods such as KNN Imputer can significantly improve model performance
Optimization of Player Experience and Enemy AI using A* Algorithm in Game
No full textThe gaming industry is rapidly evolving, where engaging and challenging gameplay has become a key factor in a game's success. Effective enemy intelligence can enhance challenges and enrich the player experience. This study aims to improve the player experience and enemy intelligence in the game Galang the EcoRescue through the implementation of the A-star (A*) algorithm. A* is a pathfinding algorithm that uses distance estimation to find the shortest path to a target by utilizing a heuristic function. This game was developed using the Unity Engine, with the implementation of the A* algorithm to determine enemy movements and adapt their behavior according to the game’s situation. Testing was conducted to ensure improvements in both the player experience and enemy intelligence. The results of the study show that the A* algorithm successfully enhanced enemy intelligence by creating more realistic and adaptive movements in response to the player, ultimately providing more dynamic challenges and improving overall gameplay quality. This study utilized the Game Development Life Cycle (GDLC) method, covering the stages of initiation, pre-production, production, testing, beta, and release. The A* algorithm has proven to optimize the player experience in Galang the EcoRescue
Optimizing Marketplace Registration Page Design with Predictive Heatmap Analysis
No full textOptimizing marketplace registration pages is crucial for improving user experience and conversion rates. This study evaluates the design of registration pages for four leading Indonesian marketplaces Tokopedia, Shopee, Blibli, and Lazada—using Predictive Heatmaps from UX Pilot alongside Heuristic Evaluation and Gestalt Principles. The analysis identifies key usability issues, such as distractions from branding elements, inconsistent visual hierarchy, and a lack of real-time validation and feedback mechanisms. Findings indicate that while branding elements effectively capture user attention, they often divert focus from essential features, a trend observed not only in these marketplaces but also in broader UI design contexts. such as Call-to-Action (CTA) buttons and registration forms. Shopee and Lazada successfully utilize high-contrast CTA buttons to direct user interaction, whereas Tokopedia and Blibli suffer from visual distractions caused by mascots and unnecessary decorative elements. Heatmap results also reveal inconsistent grouping of interface components, reducing page efficiency. To enhance user experience and conversion rates, recommendations include improving CTA button visibility through contrasting colors and strategic placement, minimizing decorative distractions, and implementing real-time validation and feedback. The application of Gestalt Principles further aids in optimizing interface organization by grouping related elements more effectively. This study underscores the importance of a structured design approach incorporating heuristic and predictive analytics to enhance the usability of online registration pages. Future research may explore the impact of interactive elements and A/B testing in refining registration interfaces
A Business Intelligence: Enhancing Apache Superset Capabilities in PBB-P2 Receivables Monitoring: -
No full textPBB-P2 Tax Revenue plays an essential role in regional finance, but managing receivables and analyzing taxpayer compliance levels still face many challenges. Business Intelligence (BI) technologies such as Apache Superset are often used for interactive data visualization. Still, they have limitations in advanced analysis, especially the application of machine learning algorithms such as K-Means for data clustering. This research aims to overcome the limitations of Apache Superset by developing an external application-based solution using the Java programming language and the SMILE library. This application is designed to cluster the level of taxpayer compliance in a batch process, with the results stored in the MySQL database. The clustered data is then visualized using Apache Superset. The results show that integrating these external applications can improve the efficiency of data analysis by utilizing more complex clustering algorithms. Visualization of clustering results also allows for more effective management of PBB-P2. This approach not only expands the capabilities of Apache Superset but also contributes to supporting data-driven tax revenue optimization strategies. This research opens up further opportunities for the integration of BI tools with machine learning algorithms in monitoring and managing complex data in the tax secto
Comparison of Xgboost, Random Forest and Logistic Regression Algorithms in Stroke Disease Classification
No full textStroke remains a critical global health concern, ranking as the second leading cause of mortality and third cause of disability worldwide. Early detection and accurate classification of stroke risk could significantly improve patient outcomes through timely interventions. This research evaluates and compares the performance of three machine learning algorithms—XGBoost, Random Forest, and Logistic Regression—for stroke disease classification using a dataset of 5,110 patient records with 12 attributes including demographic, lifestyle, and health factors. Due to significant data imbalance between stroke and non-stroke cases, Synthetic Minority Over-sampling Technique (SMOTE) was applied to enhance model performance. Comprehensive evaluation metrics including accuracy, precision, recall, and F1-score were utilized to assess each algorithm's effectiveness. Results demonstrate that XGBoost achieved superior performance with 95% accuracy, followed by Random Forest at 94% and Logistic Regression at 82%. Feature importance analysis identified age, average blood glucose level, and history of heart disease as the most significant predictors for stroke diagnosis. This study contributes to the advancement of clinical decision support systems by highlighting the effectiveness of ensemble learning approaches for stroke prediction, potentially enabling earlier interventions and improved patient management. These findings suggest that integration of machine learning tools in clinical settings could enhance stroke risk assessment, though further validation with diverse patient populations is recommended for broader implementation
Comparative Study of Forecasting Models for Smart Campus Air
No full textAir quality monitoring has become increasingly critical in urban environments, especially in densely populated smart campuses situated in tropical regions. This study presents a comparative evaluation of three predictive models CNN-GRU, LSTM, and Random Forest, for forecasting air pollution levels, specifically particulate matter concentration (PM), using real-time sensor data. The data were collected from an IoT-based monitoring system built with NodeMCU ESP8266 devices deployed on campus. Each model was trained and evaluated using performance metrics including the coefficient of determination (R²), Mean Absolute Error (MAE), and Root Mean Square Error (RMSE). The results indicate that the Random Forest model achieved the highest predictive accuracy with R² = 0.9073, MAE = 123.31, and RMSE = 274.45, outperforming both LSTM (R² = 0.8341) and CNN-GRU (R² = 0.8714). The hybrid CNN-GRU model, although capable of capturing both spatial and temporal dependencies, required larger data volumes and longer training times. The LSTM model, while effective in modeling time-series data, demonstrated a tendency to overfit when data was limited. This study highlights the practical advantages of Random Forest in modeling complex environmental data under limited resource constraints, while also demonstrating the potential of hybrid deep learning architectures. These findings contribute to the development of efficient air quality prediction systems that support health-conscious decision-making and environmental management strategies in tropical innovative campus environments
Hyperparameter Optimization with MobileNet Architecture and VGG Architecture for Urban Traffic Density Classification Using Bali Camera Image Data
No full textTraffic congestion in urban areas is a critical issue, particularly in densely populated regions such as Bali. This study addresses the challenge by implementing a Convolutional Neural Network (CNN) method to classify traffic density levels based on images captured by road surveillance cameras. The primary focus of this research is hyperparameter optimization to enhance the model's performance in classifying traffic conditions. Various combinations of hyperparameters—such as the number of neurons in the dense layer, dropout rate, learning rate, batch size, and number of epochs—were tested on two popular CNN architectures: MobileNet and VGG16. MobileNet offers lightweight computing, while VGG16 provides strong feature extraction capabilities, albeit with higher computational resource demands. Quantitative results show that after hyperparameter tuning, the MobileNet architecture achieved an accuracy of 96.94% and an F1 score of 0.969, while the VGG16 architecture achieved an accuracy of 97.22% and an F1 score of 0.972 in traffic density classification. These findings confirm that hyperparameter optimization can significantly improve classification accuracy. The scientific contribution of this research lies in the structured approach to CNN hyperparameter optimization and the demonstration that this process directly impacts the enhancement of model performance in traffic image classification tasks. This study offers valuable insights for the development of intelligent traffic management systems, especially in urban areas with limited resources