Dilara Guluzada
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Dilara Guluzada has a total of 4 co-authors. Here are the top co-authors by number of shared publications:
- Artughrul Gayibov
Baku Engineering University
1 shared publication - Esmira Mustafayeva
Department of Cybersecurity and Computer Engineering, Faculty of Information and Computer Technologies, Baku Engineering University, Hasan Aliyev str., 120, Khirdalan AZ0101, Azerbaijan
1 shared publication - Kamala Aliyeva
Baku Engineering University
1 shared publication - Vagif Gasimov
Baku Engineering University
1 shared publication
Latest Publications
- Open Access
- 30 April 2026
- Published by Private Company Technology Center in Eastern-European Journal of Enterprise Technologies
- Vol. 2 (2 (140)) , 52-64
Evaluation of a machine learning-assisted interactive evolutionary non-dominated sorting genetic algorithm -II framework for hyperparameter optimization of U-Net in agricultural land segmentation
The object of the study is the hyperparameter configuration space of the U-Net architecture for agricultural land segmentation from Sentinel-2 satellite imagery. The problem being solved is the excessive cost of multi-objective hyperparameter optimization, because non-dominated sorting in the non-dominated sorting genetic algorithm II (NSGA-II), with complexity O(MN2), becomes a bottleneck for deep segmentation models. To address this problem, an interactive evolutionary non-dominated sorting genetic algorithm II (IENSGA-II) framework is evaluated, in which a logistic regression classifier is trained on hyperparameter vectors and Pareto ranks from initial NSGA-II generations, then used to predict ranks in subsequent generations instead of full sorting. Unlike existing surrogate-assisted approaches, this work predicts Pareto ranks without additional model evaluations. On the panoptic agricultural satellite time series (PASTIS) benchmark, the framework reduced execution time by 20.07%, 16.39%, and 38.80% for 5, 10, and 15 generations, and in the 10-generation setting improved validation criteria, reaching an area under the receiver operating characteristic curve (AUC) of 0.9072 versus 0.9004 and validation loss of 0.6057 versus 0.6212. These results were achieved because the method accelerates selection rather than replacing model evaluation, while AUC-based tie-breaking preserves preference for more accurate solutions among candidates with same predicted rank. Effectiveness stems from a regular relationship between hyperparameters and Pareto ranks in early evolutionary data. In practice, the method is used in resource-constrained multi-objective learning when initial generations provide representative data for rank prediction
- 30 April 2026
- Published by Private Company Technology Center in Eastern-European Journal of Enterprise Technologies
- Vol. 2 (2 (140)) , 52-64
Open AccessEvaluation of a machine learning-assisted interactive evolutionary non-dominated sorting genetic algorithm -II framework for hyperparameter optimization of U-Net in agricultural land segmentation
The object of the study is the hyperparameter configuration space of the U-Net architecture for agricultural land segmentation from Sentinel-2 satellite imagery. The problem being solved is the excessive cost of multi-objective hyperparameter optimization, because non-dominated sorting in the non-dominated sorting genetic algorithm II (NSGA-II), with complexity O(MN2), becomes a bottleneck for deep segmentation models. To address this problem, an interactive evolutionary non-dominated sorting genetic algorithm II (IENSGA-II) framework is evaluated, in which a logistic regression classifier is trained on hyperparameter vectors and Pareto ranks from initial NSGA-II generations, then used to predict ranks in subsequent generations instead of full sorting. Unlike existing surrogate-assisted approaches, this work predicts Pareto ranks without additional model evaluations. On the panoptic agricultural satellite time series (PASTIS) benchmark, the framework reduced execution time by 20.07%, 16.39%, and 38.80% for 5, 10, and 15 generations, and in the 10-generation setting improved validation criteria, reaching an area under the receiver operating characteristic curve (AUC) of 0.9072 versus 0.9004 and validation loss of 0.6057 versus 0.6212. These results were achieved because the method accelerates selection rather than replacing model evaluation, while AUC-based tie-breaking preserves preference for more accurate solutions among candidates with same predicted rank. Effectiveness stems from a regular relationship between hyperparameters and Pareto ranks in early evolutionary data. In practice, the method is used in resource-constrained multi-objective learning when initial generations provide representative data for rank prediction