Filiz Ersöz

Despite the worldwide move towards smart meters, mechanical water meters are still used extensively in residential and municipal networks. Upgrading them with computer vision avoids costly infrastructure replacement, yet most approaches send data to cloud servers for processing. This paper introduces an end-to-end on-device solution for recognizing water meter digits on the ESP32-CAM board, which costs below $15 and features only 520 KB SRAM. A MobileNetV2-based digit classifier is quantized to INT8. Three firmware improvements extend the inference loop: i) a temporal check confirming a reading only when three consecutive detections match, ii) on-device anomaly detection without server support, and iii) adaptive sampling between 3 s during water usage and 60 s when idle. A Flask-based internet of things (IoT) dashboard named MeterWatch provides live monitoring. The model achieves 2 ms latency and 235.6 KB peak memory. Ona553-imagetest subset from a custom 6,155-image dataset with 13,470 digit annotations it reaches 99.60% cell-level accuracy including background, 96.89% digit-only accuracy and 87.70% exact five-digit reading accuracy. To the best of our knowledge, this is the first system combining on-device INT8 deployment, water meter digit recognition and intelligent firmware behaviour on the ESP32-CAM. Evaluations used a single meter brand under fixed flash illumination.

This study examines whether order-line product and transaction data can be used to classify items into female and male segments in fashion e-commerce. Sales amount, discount amount, size, color group, and product type were used as predictors, and SVM, KNN, Decision Tree, and Random Forest algorithms were compared. Random Forest delivered the best overall performance. To move beyond predictive accuracy, SHAP, LIME, and DiCE were employed to investigate how the model formed its decisions. The findings show that the model relies primarily on financial variables, especially sales amount and discount amount, while product type and other categorical variables provide secondary but still meaningful contributions. Local explanations further indicate that the decision structure is not fully symmetric across female and male classes and that several predictions depend strongly on sales amount thresholds. By combining order-line classification with multi-layered explainability, the study contributes a methodologically grounded perspective on data-driven segment differentiation in fashion e-commerce.

Cost estimation is crucial for competitiveness in capital-intensive industries. This study compares Multiple Linear Regression and Artificial Neural Networks for forecasting crude steel costs using Turkish industry data from 2008–2019 across SPSS, Python, Minitab, and EViews. Results show that MLR achieves superior predictive accuracy (R2 > 0.99), particularly in SPSS and Python, while liquid steel cost is the most influential driver of total production expenses. Although ANN captures nonlinear dynamics, its error rates remain higher than regression models. Overall, findings highlight the robustness of MLR in structured datasets and the potential of ANN for nonlinear or incomplete data, offering practical insights for cost minimization, resource allocation, and competitiveness in the steel sector.

Crime prediction has become a valuable tool for enhancing predictive policing, enabling law enforcement agencies to allocate resources more effectively and implement proactive crime prevention strategies, particularly in high-crime areas. The use of artificial intelligence (AI) has revolutionized this field by analyzing vast amounts of data to identify patterns and anticipate criminal activities with unprecedented accuracy. This paper aims to review the literature on AI-based crime prediction, analyzing 142 studies that focus on crimes against individuals, society, and property. Despite the promising potential of AI in crime prediction, significant challenges remain, particularly regarding the trustworthiness of AI systems, which is essential for their social acceptance. To address these issues, this review explores the explainability of AI-based prediction models, with a specific focus on the role of explainable AI (XAI). The findings highlight the importance of XAI in building trust in these models by offering more transparent and interpretable insights into how AI systems make decisions. However, the review also reveals that the integration of XAI remains underdeveloped in the current literature. By improving the transparency of AI systems, XAI has the potential to lead to more accurate, trustworthy, and fair crime predictions, ultimately facilitating more effective and equitable crime prevention efforts.

Urban seismic risk assessments in Istanbul have predominantly focused on district level loss estimates, even though mitigation and emergency response decisions are implemented at much finer administrative units. This study develops a neighborhood-based classification of earthquake risk for all 959 neighborhoods under a deterministic Mw 7.5 scenario. The analysis relies on the official Istanbul Earthquake Loss Estimation Update dataset prepared by Istanbul Metropolitan Municipality in cooperation with the Kandilli Observatory. Eight scenario-based outcome indicators, including four structural damage and four human impact variables, are first transformed using Yeo-Johnson and Min-Max scaling and then reduced through Principal Component Analysis, which explains 96.3 percent of the total variance in two components. Within this reduced space, K-Means, K-Medoids, Gaussian Mixture Models, Spectral Clustering, and HDBSCAN are systematically compared. Model selection is guided by internal validation criteria, the Gap Statistic, and bootstrap stability analysis. Based on this combined assessment, K-Medoids with k equal to 2 emerges as the most parsimonious and stable clustering solution. The resulting High Impact and Low Impact profiles show statistically significant differences across all indicators and remain highly consistent across 300 bootstrap resamples, with a mean Adjusted Rand Index of 0.976. The identified medoid neighborhoods provide concrete reference cases for targeted planning interventions. Spatially, higher impact areas are concentrated along the Marmara coastal corridor and older urban cores, whereas lower impact neighborhoods are more common in northern districts. By converting detailed scenario outputs into stable neighborhood level risk categories, the study provides a structured basis for prioritizing mitigation investments, preparedness actions, and emergency response planning at the local scale.