Jurnal Sistem Informasi dan Komputer Terapan Indonesia (JSIKTI)

Classifying UKT Fee Relief Eligibility Using K Nearest Neighbors Algorithm

I Wayan Kintara Anggara Putra — 2025-11-16

This research develops a K-Nearest Neighbors (KNN)-based classification model to determine the eligibility of students for Tuition Assistance (UKT) based on socio-economic factors, including parental income, family size, parental occupation, number of dependents, and housing conditions. The goal is to automate the process of identifying students eligible for financial aid, enhancing both the efficiency and fairness in resource allocation. The model was trained using a dataset consisting of both categorical and numerical features, with the target variable being binary: "Eligible" (1) or "Not Eligible" (0) for UKT relief. The KNN model achieved an overall accuracy of 92%, with strong performance in predicting the "Eligible" class. However, the "Not Eligible" class showed lower performance, particularly in terms of recall and F1-score, suggesting the presence of class imbalance. To address this issue, techniques such as class balancing, resampling, or adjusting KNN parameters are suggested to improve the model's ability to correctly classify minority instances. Additionally, exploring ensemble methods like Random Forest or XGBoost may provide more robust results. This study highlights the importance of addressing class imbalance and using appropriate evaluation metrics beyond accuracy when building classification models for imbalanced datasets.

Decision Tree Model for Predicting Ethereum Price Movements Based on Trends

I Dewa Ayu Sri Murdhani — 2025-11-16

This research investigates the application of a Decision Tree model for predicting Ethereum price movements using historical trend data. The dataset includes key attributes such as open, high, low, close prices, and trading volume, offering insights into market dynamics. The research emphasizes preprocessing and feature engineering techniques, including normalization and the introduction of derived metrics like moving averages and Relative Strength Index (RSI). Despite the model's simplicity and interpretability, it achieved an accuracy of 49.10%, indicating limited effectiveness in capturing non-linear relationships in volatile cryptocurrency markets. Analysis reveals challenges in distinguishing price trends and handling data imbalances, leading to suboptimal performance. These findings highlight the complexities of financial prediction and underscore the need for advanced machine learning methods. Future work should explore ensemble models, richer datasets incorporating sentiment analysis, and resampling techniques to improve robustness and predictive accuracy. This research contributes to the growing literature on machine learning applications in cryptocurrency analytics.

Efficient Wine Quality Prediction and Classification Using LightGBM Model

Muslimin B — 2025-11-16

This study develops an efficient machine learning model using the Light Gradient Boosting Machine (LightGBM) algorithm to predict and classify wine quality based on physicochemical properties. The dataset used in this research consists of multiple chemical attributes, including alcohol content, acidity levels, sulphates, and phenolic compounds, which collectively influence wine quality. The preprocessing stage involved data cleaning, outlier treatment, feature scaling, and handling class imbalance using the Synthetic Minority Oversampling Technique (SMOTE). Feature selection was conducted using mutual information and recursive feature elimination to identify the most influential predictors. The optimized LightGBM model achieved superior performance with 100% accuracy, precision, recall, and F1-score across all quality classes, outperforming traditional algorithms such as Random Forest, SVM, and Logistic Regression. Feature importance analysis revealed that Proline, Flavanoids, and Magnesium were the most significant attributes contributing to wine classification. These findings demonstrate that LightGBM is a robust and scalable solution for wine quality prediction, offering an efficient, data-driven alternative to traditional sensory evaluations. The proposed model can enhance quality control processes in the wine industry by providing accurate and interpretable insights into the chemical determinants of wine quality.

Naive Bayes Classifier for Accurate Diabetes Diagnosis and Analysis

Lynn Htet Aung — 2025-11-16

Diabetes mellitus is a chronic metabolic disorder with rising global prevalence, necessitating early and accurate diagnostic tools to mitigate complications. This study investigates the Naive Bayes classifier's efficacy for diabetes diagnosis, leveraging a dataset of 768 patient records encompassing clinical and demographic attributes, such as glucose levels, BMI, and insulin. Data preprocessing steps, including imputation, scaling, and normalization, ensure data quality, while feature selection identifies key predictors to enhance model performance. The classifier achieved an accuracy of 77%, with a weighted F1-score of 0.77, demonstrating robust performance for the "Not Worthy" class but moderate results for the "Worthy" class due to class imbalance and overlapping features. Ensemble methods, such as bagging and boosting, were explored to address these challenges, further improving robustness and recall. The study highlights the Naive Bayes classifier as a cost-effective, computationally efficient tool for real-time diabetes detection, with potential for deployment in resource-limited healthcare settings. Future research should focus on class balancing, advanced feature engineering, and validation on larger, diverse datasets to enhance diagnostic reliability and scalability.

Time Series Prediction of Doge Coin Prices Using LSTM Networks

Aniek Suryanti Kusuma — 2025-11-16

This research explores the application of Long Short-Term Memory (LSTM) networks for predicting Dogecoin prices, addressing the challenges of cryptocurrency market volatility and non-linearity. A historical dataset spanning November 2017 to the present, including features such as opening and closing prices, daily highs and lows, and trading volume, was used for model development. Data preprocessing involved handling missing values, normalization, and structuring the data into a supervised learning format. The LSTM model was designed with optimized hyperparameters, trained using the Adam optimizer, and evaluated against metrics such as Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE). Benchmarking with traditional models like ARIMA and SVR demonstrated the LSTM model's superior performance in capturing temporal dependencies and adapting to high volatility. Despite its robust performance, the study highlights limitations, including the exclusion of external factors like market sentiment and a dataset limited to specific timeframes. Future research could integrate broader datasets and advanced features to enhance model precision. This work contributes to the field of cryptocurrency forecasting, offering insights for traders, investors, and researchers navigating volatile markets.