Vol. 1, No. 1, [Month] 2025

Wireless communication systems have been experiencing a dramatic increase in data traffic volume due to the introduction of new services and applications in every up-to-date generation of mobile wireless communication technology. In 5G and beyond communication systems, the emerging services and use cases (mission critical communication, massive Internet of Things, enhanced mobile broadband, and ultra-reliable communication) are envisaged to transform the digital society in the industrial sector. The heterogeneous cellular network (HCN) technology is suitable for handling the data traffic and fading, thus resulting in a high network coverage capacity and data rate of the system. However, deployment of numerous base stations in a cellular network presents crucial challenges owing to their significantly high power consumption. To this end, this paper proposes a technique comprising a topology of distributed access points (APs) in the network coverage area for wireless communications to improve the energy efficiency by minimizing power consumption. The AP operations are controlled by a powerful central processing unit (CPU). The simulation results demonstrated that the proposed network topology significantly improves energy efficiency (approximately 60%) compared to the conventional heterogeneous cellular network, thus proving itself to be relatively environment friendly.

Credit card fraud detection remains challenging because fraudulent transactions represent less than 0.2% of all transactions, creating severe class imbalance. Existing studies often apply feature selection and SMOTE without clearly specifying whether these operations are performed independently within cross-validation folds, which creates a risk of data leakage and inflated performance estimates. This study proposes a fold-quarantined iterative Knowledge Discovery in Databases (KDD) pipeline in which feature selection, SMOTE, feature scaling, model training, and decision-threshold calibration are performed exclusively on the training data of each fold. The pipeline is evaluated with stratified five-fold cross-validation on the Kaggle Credit Card Fraud Detection dataset (284,807 transactions, 492 of them fraudulent, 0.172%) using a Multilayer Perceptron (MLP) trained with focal loss, Logistic Regression, Decision Tree, and Random Forest. Random Forest achieved the best overall performance (F1 = 0.843), followed by the focal-loss MLP (F1 = 0.815), whereas the Decision Tree and Logistic Regression performed considerably worse. The results show that model performance under leakage-controlled validation varies considerably across models, and the proposed pipeline offers a transparent, reproducible protocol for evaluating fraud-detection models under severe class imbalance.

Keywords: Credit card fraud detection; data leakage; cross-validation; SMOTE; threshold calibration; focal loss; class imbalance.

The growing adoption of automated information extraction in modern enterprises has increased the importance of accurately segmenting tables and their columns from document images. This task remains challenging due to the coexistence of tables with textual and graphical content, as well as substantial variations in table dimensions, layouts, and boundary structures. To address these challenges, we introduce ResForTable, a U-Net-based architecture featuring a pretrained group-normalized ResNet-50 encoder and two transformer-enhanced decoder branches dedicated to table and column segmentation. Evaluation on the Marmot dataset yields DICE scores of 0.9352 for table segmentation and 0.9045 for column segmentation. ResForTable consistently surpasses TransTab across the evaluated metrics while reducing computational requirements by 42.7% in terms of FLOPs compared with a comparable VGG-19-based architecture. Further analysis identifies a persistent difference between overlap-based DICE and boundary-based ASSD measurements across both segmentation branches. Based on Weyl’s Tube Formula, the DICE deficit is modeled as a function of ASSD and the mask perimeter-to-area ratio (P/A). The statistically significant relationship between DICE deficit and P/A indicates that DICE can penalize geometric shape even when segmentation boundaries are accurately positioned. These findings emphasize the need to employ complementary evaluation metrics for more reliable assessment of table and column segmentation.

Keywords: Table Segmentation; Column Segmentation; Document Image Analysis; ResNet-Transformer; Deep Learning.