16 July 2025 in International Journal of Engineering, Science and Information Technology
The issue of the high number of students dropping out of college is a major concern in higher education, especially in the smart campus ecosystem. This research aims to design a prediction system for students who are at risk of dropping out by integrating hybrid feature selection methods and ensemble learning that leverage academic data and students' digital footprints. The initial process of model development involves data cleaning and the selection of important features through a combination approach using filter-based methods (mutual information) and recursive feature elimination. A classification model is then designed using the XGBoost and Random Forest algorithms. The testing was conducted using a secondary dataset that included variables such as participation in discussions, attendance rates, interaction with learning materials, and academic achievement. The results of testing with the XGBoost model showed a satisfactory accuracy level, with an F1 score of 0.77 and a ROC AUC of 0.89. The confusion matrix recorded 67 correct predictions for students who graduated and 17 correct predictions for students who dropped out, with a total of 12 misclassifications. These findings suggest that the combination of hybrid feature selection strategies and XGBoost can produce sufficiently accurate predictions of student dropouts and has the potential to be utilized as an early warning system in the governance of a more flexible and responsive smart campus.
22 October 2026 in Eastern-European Journal of Enterprise Technologies
The object of the study is the dataset- and classifier-dependent behavior of a frozen Latin hypercube sampling and nearest-neighbor rebalancing procedure for imbalanced tabular classification. The problem addressed was the absence of practical evidence showing when rebalancing improves, leaves unchanged, or degrades performance across datasets and classifier families. The procedure was evaluated on 20 tabular datasets using five deterministic seeds, duplicate-aware cross-validation, eight sampling conditions, and four classifiers. Relative to synthetic minority over-sampling, it increased dataset-balanced macro-averaged F1 score from 0.7717 to 0.7761, produced 13 dataset-level wins and 7 losses, and yielded a Wilcoxon p-value of 0.0441 with a paired bootstrap 95% confidence interval of [0.00104, 0.00788]. The mean dataset-level difference was +0.004389. The largest classifier-specific gain in the dataset-balanced summaries occurred with distance-weighted seven-nearest neighbors (+0.011944); gains for extreme gradient boosting, random forest, and multilayer perceptron were small. Six dataset characteristics showed no significant monotonic association with gains after Holm correction. This does not establish causal effects of local geometry or classifier induction. The distinctive result is a reproducible classification of benefit, neutral, and loss regimes supported by dataset-level inference, classifier contrasts, minority-class outcomes, and runtime analysis. Mean total runtime was 2.89 times that of the SMOTE comparator. The findings guide comparative validation when local component preservation, controlled synthetic generation, and neighborhood-compatible classification are important
30 September 2026 in SISTEMASI
SSSyahrial Jeremia Sinaga
VSVivaldi Adventus Simangunsong
Vivaldi Adventus Simangunsong
18 January 2021 in Journal of Informatics and Telecommunication Engineering
Al-Quran is the basic reason someone should understand the rules of Arabic. One of the basic rules is knowing the jar type which we generally often encounter in Arabic, but we do not understand that this letter has its own duties and functions. In this study, samples of Jar letters were used as many as 7 jar letter patterns which are generally often encountered in the Al-Quran. The purpose of this research is to build a system that can recognize Jar letters using the Pierce Similarity method and performs the performance on the algorithm. The research method used is the theory of pattern recognition in image processing with 2 processes, namely the Training Process and the Testing Process. The value of each letter pattern obtained in the Training Process will be the weight benchmark for the Testing Process, so that we can measure the performance level of Algortima Pierce Similarity in detecting the Jar letter pattern. The results can vary for each letter pattern ranging from 60% to 80%.