Journal of Systems Engineering and Information Technology (JOSEIT)

Enhancing News Recommendations with Deep Reinforcement Learning and Dynamic Action Masking

Dong Sang-hong, Ahn Jun-soo — Mon, 28 Apr 2025 10:56:40 +0000

A news recommender system is crucial for the transmission of news in new media. A deep reinforcement learning-based recommender system is suggested to integrate the characterization capabilities of neural networks with the strategic selection capabilities of reinforcement learning to enhance news recommendation efficacy. Dynamic action masks enhance the capacity to assess short-term interests of users. An optimized caching mechanism improves the efficiency of the experience cache, and a reward design characterized by region masking accelerates model training, thereby enhancing the performance of the recommender system for news recommendations. Experimental results indicate that the recommendation accuracy of the proposed model on the news dataset is on par with that of prevalent neural network recommendation techniques and surpasses existing state-of-the-art algorithms in ranking performance

Comparison and Optimization of Parallel Clustering Algorithms for Chinese A-Share Stock Segmentation Based on Financial Indicators

Hai Mo, Niu Yihan, Zhang Yuejin — Wed, 30 Apr 2025 00:00:00 +0000

This study presents a novel application of parallel clustering algorithms to segment stocks in the Chinese A-share market based on financial indicators. Using the Hadoop platform and Mahout software library, we implemented and compared the performance of the K-means and fuzzy K-means algorithms across five distance measures: Euclidean, squared Euclidean, Manhattan, cosine, and Tanimoto. The analysis utilized 15 financial indicators from 2,544 listed companies to reflect profitability, solvency, growth capability, asset management quality, and shareholder profitability. The experimental results demonstrate that for stock financial data clustering, the K-means algorithm with Tanimoto distance yields optimal execution efficiency and clustering quality, whereas the fuzzy K-means algorithm performs best with squared Euclidean distance. However, the K-means algorithm proved to be more effective overall, successfully categorizing 1,483 stocks into 26 meaningful segments compared to only 511 stocks in 27 segments using fuzzy K-means. The resulting stock segmentation framework divides the market into eight comprehensive categories based on investment value and security, thereby providing investors with practical guidance for stock selection. Our approach enables investors to understand the fundamental characteristics of each stock segment, discern their distinctive features, and identify undervalued stocks with appreciative potential. This study represents the first application of parallel big data clustering algorithms to segment the entire Chinese A-share market, offering significant practical value for investment decision-making.

Multi-Class CNN Models for Banana Ripeness Classification

Rafaela S. Francisco, Gabriel de S. G. Pedroso, Thiago M. Ventura — Wed, 30 Apr 2025 00:00:00 +0000

This study develops and evaluates Convolutional Neural Network (CNN) models for classifying banana maturity stages using images, thereby addressing a significant challenge in the banana supply chain. The banana industry represents a major agricultural sector worldwide, with Brazil exporting 56.2 thousand tons in 2023. Accurate maturity classification is essential for optimizing harvest timing, reducing postharvest losses, and extending shelf life. We utilized a public Brazilian dataset of 1,000 images of Prata Catarina banana categorized into eight ripening stages based on peel coloration standards established by the Brazilian Program for Horticulture Modernization. The images were preprocessed to a standardized 200 × 200-pixel resolution, and we evaluated the effectiveness of the data augmentation techniques, including horizontal flip, vertical flip, rotation, and zoom. Our CNN architecture consisted of five convolutional blocks with a dropout layer prior to flattening. We conducted six experiments to compare three classification scenarios (eight, five, and two ripeness classes) with and without data augmentation. Our findings demonstrate that CNN models can effectively classify banana ripeness, with performance improving significantly as classification granularity decreases. The best-performing model achieved 89.5% accuracy, 87.2% precision, and 89.6% recall when classifying bananas into two categories.

Machine Learning Models for Air Pollution Health Risk Assessment

Lipatova A.V, Potapchenko T.D — Wed, 30 Apr 2025 00:00:00 +0000

This study explored the application of machine learning (ML) models and artificial neural networks (ANNs) in the assessment of public health concerns associated with air pollution. Utilizing a dataset comprising over 12,000 records from India and Nepal, encompassing both quantitative measurements and visual data, several classification models have been constructed and evaluated to predict air quality index (AQI) categories indicative of varying health risk levels. The implemented models comprise a decision tree (DT), support vector machine (SVM), random forest (RF), XGBoost, and deep neural networks (both convolutional and recurrent). The methodology entailed data preprocessing, feature significance analysis, and model assessment using accuracy metrics and ROC curves. The findings revealed a high classification accuracy across all models (>90%), with ensemble-based methods demonstrating enhanced performance. XGBoost attained superior accuracy with optimal resource efficiency; however, artificial neural network (ANN) models, particularly long short-term memory (LSTM), obtained accuracy levels of 98% by the 15th training epoch. Feature significance analysis revealed that AQI, PM2.5, and PM10 were the primary predictors of health risk categorization. Correlation analysis demonstrated robust associations between particulate matter measures (PM2.5, PM10), underscoring their significance in air quality evaluation. This study proposes a methodological framework for automating risk assessment procedures using machine learning approaches to facilitate more effective environmental health monitoring. The findings suggest that ensemble models offer an optimal balance between precision and computing efficiency for real-time air quality classification systems with potential applications in early warning systems and public health intervention techniques.

Analysis of an Adaptive E-Learning System with the Adjustment of the Felder-Silverman Model in Moodle

Heni Jusuf, Andiani — Wed, 30 Apr 2025 00:00:00 +0000

In the digital era, adaptive e-learning has become essential in addressing students’ diverse learning preferences. This study aims to develop an adaptive e-learning system that integrates the Felder-Silverman Learning Style model (FSLSM) into Moodle using fuzzy logic and case-based reasoning. The system extracts behavioral attributes from student activity logs and classifies learning styles into four dimensions: processing, perception, input, and understanding. The experimental evaluation, conducted with and without substitution of the (ILS) questionnaire values, demonstrated varying levels of accuracy. Accuracy improved with ILS substitution as follows: processing (82.86%), perception (80.00%), input (80.00%), and understanding (74.29%). Without ILS substitution, the accuracies were as follows: processing (80.00%), perception (80.00%), input (74.29%), and understanding (62.86%). These findings confirm the system’s potential to support personalized learning by accurately identifying learning styles.