Muhammed Abdulhamid Karabiyik
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Muhammed Abdulhamid Karabiyik has a total of 6 co-authors. Here are the top co-authors by number of shared publications:
- Bahaeddin Turkoglu
Department of Artificial Intelligence and Data Engineering, Ankara University, Ankara, Trkiye
4 shared publications - Tunc Asuroglu
Tampere University
4 shared publications - Asım Sinan Yüksel
Süleyman Demirel University
3 shared publications - Fatma Gülşah Tan
Süleyman Demirel Üniversitesi
3 shared publications - Huseyin Eldem
Department of Software Development, Karamanoğlu Mehmetbey University, Karaman 70100, Turkey
1 shared publication - Mustafa Boğa
Niğde Ömer Halisdemir University, Bor Vocational School, TR51700, Bor, Niğde, Türkiye
1 shared publication
Latest Publications
- Open Access
SC-EfficientNet: Preventing Attention Collapse in Consistency-Regularized Skin Lesion Classification
Attention-based consistency regularization aims to focus skin lesion classifiers on lesions, but whether the attention behaves as intended is rarely verified. The main contribution is the analysis and prevention of attention collapse in Kullback–Leibler (KL) consistency training, not improved classification performance. In SC-EfficientNet (Supervised-Consistency EfficientNet), an attention-masked prediction branch is supervised with the class labels and regularized toward the original prediction by a Kullback–Leibler term, and its spatial mask is constrained by area and binarization priors. Models were evaluated with lesion-level five-fold cross-validation on ISIC 2018 Task 3 (three seeds, EfficientNet-B0/B4) and on the official test set. Trained only through the consistency term, the mask collapsed to a nearly constant map in every run, also with a tenfold larger consistency weight, a shared head, or zero-initialized attention. SC-EfficientNet prevented this collapse, produced spatially selective masks, and raised agreement with Grad-CAM from r = 0.04–0.08 to 0.31–0.44. A matched headless control explained most of the B4 gain over a standard baseline; relative to this control, SC-EfficientNet improved B0 macro-F1 by 2.16 ± 0.31 points but did not change B4 performance (85.12% ± 0.10% out-of-fold accuracy), and no variant improved robustness to synthetic artifacts. Melanoma sensitivity under the default decision rule remained low (SC-EfficientNet-B4: 53.5% out of fold, 55.2% on the test set). Matched controls and mask diagnostics are needed before attributing gains to attention.
Open AccessSC-EfficientNet: Preventing Attention Collapse in Consistency-Regularized Skin Lesion Classification
Attention-based consistency regularization aims to focus skin lesion classifiers on lesions, but whether the attention behaves as intended is rarely verified. The main contribution is the analysis and prevention of attention collapse in Kullback–Leibler (KL) consistency training, not improved classification performance. In SC-EfficientNet (Supervised-Consistency EfficientNet), an attention-masked prediction branch is supervised with the class labels and regularized toward the original prediction by a Kullback–Leibler term, and its spatial mask is constrained by area and binarization priors. Models were evaluated with lesion-level five-fold cross-validation on ISIC 2018 Task 3 (three seeds, EfficientNet-B0/B4) and on the official test set. Trained only through the consistency term, the mask collapsed to a nearly constant map in every run, also with a tenfold larger consistency weight, a shared head, or zero-initialized attention. SC-EfficientNet prevented this collapse, produced spatially selective masks, and raised agreement with Grad-CAM from r = 0.04–0.08 to 0.31–0.44. A matched headless control explained most of the B4 gain over a standard baseline; relative to this control, SC-EfficientNet improved B0 macro-F1 by 2.16 ± 0.31 points but did not change B4 performance (85.12% ± 0.10% out-of-fold accuracy), and no variant improved robustness to synthetic artifacts. Melanoma sensitivity under the default decision rule remained low (SC-EfficientNet-B4: 53.5% out of fold, 55.2% on the test set). Matched controls and mask diagnostics are needed before attributing gains to attention.
- Open Access
- 31 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 187740-187757
HENOS: A Henon Map-Based Chaotic Oversampling Strategy for Imbalanced Data Classification
The effectiveness of machine learning models is profoundly influenced by the quality and distribution of training data. However, real-world datasets are often highly imbalanced, where conventional classification algorithms tend to favor the majority class, resulting in poor recognition of minority instances. To address this challenge, we propose HENOS, a novel oversampling method that leverages the nonlinear dynamics of the Henon map to synthesize minority class samples in a more structurally diverse and boundary-aware manner. Unlike traditional interpolation-based techniques such as SMOTE, which often fail to capture complex data distributions, HENOS introduces a deterministic yet chaotic mechanism to generate synthetic instances that preserve local data characteristics while enhancing class separability. We conduct comprehensive experiments on 37 benchmark datasets, evaluating HENOS across three distinct classifiers AdaBoost, Naive Bayes, and Artificial Neural Networks (ANN) using Area Under the Curve (AUC) and F1-score metrics. The results, validated by the Friedman statistical test, demonstrate that HENOS consistently outperforms existing oversampling methods with statistically significant gains. These findings highlight the potential of chaos-theoretic principles in tackling data imbalance and open avenues for their integration into advanced learning systems, including deep architectures.
- 31 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 187740-187757
Open AccessHENOS: A Henon Map-Based Chaotic Oversampling Strategy for Imbalanced Data Classification
The effectiveness of machine learning models is profoundly influenced by the quality and distribution of training data. However, real-world datasets are often highly imbalanced, where conventional classification algorithms tend to favor the majority class, resulting in poor recognition of minority instances. To address this challenge, we propose HENOS, a novel oversampling method that leverages the nonlinear dynamics of the Henon map to synthesize minority class samples in a more structurally diverse and boundary-aware manner. Unlike traditional interpolation-based techniques such as SMOTE, which often fail to capture complex data distributions, HENOS introduces a deterministic yet chaotic mechanism to generate synthetic instances that preserve local data characteristics while enhancing class separability. We conduct comprehensive experiments on 37 benchmark datasets, evaluating HENOS across three distinct classifiers AdaBoost, Naive Bayes, and Artificial Neural Networks (ANN) using Area Under the Curve (AUC) and F1-score metrics. The results, validated by the Friedman statistical test, demonstrate that HENOS consistently outperforms existing oversampling methods with statistically significant gains. These findings highlight the potential of chaos-theoretic principles in tackling data imbalance and open avenues for their integration into advanced learning systems, including deep architectures.
- Open Access
- 14 October 2025
- Published by PeerJ in PeerJ Computer Science
- Vol. 11, e3257
RefactorGPT: a ChatGPT-based multi-agent framework for automated code refactoring
The rise of large language models has redefined what is computationally possible in code generation, yet their potential in systematic software refactoring remains largely untapped. This article introduces RefactorGPT, a ChatGPT-augmented sequential multi-agent framework that transforms refactoring from a monolithic, opaque process into a modular, explainable, and scalable workflow. The system orchestrates four specialized agents, Analyzer, Refactor, Refine, and Fixer, to sequentially analyse source code, apply structural refactoring, enhance code quality, and recover from potential generation errors. Unlike conventional rule-based tools or one-shot large language model (LLM) prompts, RefactorGPT leverages ChatGPT iteratively across clearly defined responsibilities, enabling controlled refactoring with functional guarantees. To evaluate the framework’s effectiveness, we constructed a curated dataset encompassing nine classical refactoring techniques across three complexity levels. RefactorGPT demonstrated consistent improvements in code modularity, readability, and structural decomposition, while maintaining computational efficiency. Notably, the system achieved full execution correctness through autonomous error recovery, showcasing robustness in practical scenarios. This study contributes a reusable blueprint for LLM-integrated refactoring systems and presents a novel application of ChatGPT not merely as a code generator, but as a cooperative agent in intelligent software refactoring. The findings reveal a path forward for embedding language models into real-world developer workflows, not as assistants, but as collaborators in code evolution.
- 14 October 2025
- Published by PeerJ in PeerJ Computer Science
- Vol. 11, e3257
Open AccessRefactorGPT: a ChatGPT-based multi-agent framework for automated code refactoring
The rise of large language models has redefined what is computationally possible in code generation, yet their potential in systematic software refactoring remains largely untapped. This article introduces RefactorGPT, a ChatGPT-augmented sequential multi-agent framework that transforms refactoring from a monolithic, opaque process into a modular, explainable, and scalable workflow. The system orchestrates four specialized agents, Analyzer, Refactor, Refine, and Fixer, to sequentially analyse source code, apply structural refactoring, enhance code quality, and recover from potential generation errors. Unlike conventional rule-based tools or one-shot large language model (LLM) prompts, RefactorGPT leverages ChatGPT iteratively across clearly defined responsibilities, enabling controlled refactoring with functional guarantees. To evaluate the framework’s effectiveness, we constructed a curated dataset encompassing nine classical refactoring techniques across three complexity levels. RefactorGPT demonstrated consistent improvements in code modularity, readability, and structural decomposition, while maintaining computational efficiency. Notably, the system achieved full execution correctness through autonomous error recovery, showcasing robustness in practical scenarios. This study contributes a reusable blueprint for LLM-integrated refactoring systems and presents a novel application of ChatGPT not merely as a code generator, but as a cooperative agent in intelligent software refactoring. The findings reveal a path forward for embedding language models into real-world developer workflows, not as assistants, but as collaborators in code evolution.
- Open Access
- 2 September 2025
- Published by PeerJ in PeerJ Computer Science
- Vol. 11, e3177
A cluster-assisted differential evolution-based hybrid oversampling method for imbalanced datasets
Class imbalance remains a significant challenge in machine learning, leading to biased models that favor the majority class while failing to accurately classify minority instances. Traditional oversampling methods, such as Synthetic Minority Over-sampling Technique (SMOTE) and its variants, often struggle with class overlap, poor decision boundary representation, and noise accumulation. To address these limitations, this study introduces ClusterDEBO, a novel hybrid oversampling method that integrates K-Means clustering with differential evolution (DE) to generate synthetic samples in a more structured and adaptive manner. The proposed method first partitions the minority class into clusters using the silhouette score to determine the optimal number of clusters. Within each cluster, DE-based mutation and crossover operations are applied to generate diverse and well-distributed synthetic samples while preserving the underlying data distribution. Additionally, a selective sampling and noise reduction mechanism is employed to filter out low-impact synthetic samples based on their contribution to classification performance. The effectiveness of ClusterDEBO is evaluated on 44 benchmark datasets using k-Nearest Neighbors (kNN), decision tree (DT), and support vector machines (SVM) as classifiers. The results demonstrate that ClusterDEBO consistently outperforms existing oversampling techniques, leading to improved class separability and enhanced classifier robustness. Moreover, statistical validation using the Friedman test confirms the significance of the improvements, ensuring that the observed gains are not due to random variations. The findings highlight the potential of cluster-assisted differential evolution as a powerful strategy for handling imbalanced datasets.
- 2 September 2025
- Published by PeerJ in PeerJ Computer Science
- Vol. 11, e3177
Open AccessA cluster-assisted differential evolution-based hybrid oversampling method for imbalanced datasets
Class imbalance remains a significant challenge in machine learning, leading to biased models that favor the majority class while failing to accurately classify minority instances. Traditional oversampling methods, such as Synthetic Minority Over-sampling Technique (SMOTE) and its variants, often struggle with class overlap, poor decision boundary representation, and noise accumulation. To address these limitations, this study introduces ClusterDEBO, a novel hybrid oversampling method that integrates K-Means clustering with differential evolution (DE) to generate synthetic samples in a more structured and adaptive manner. The proposed method first partitions the minority class into clusters using the silhouette score to determine the optimal number of clusters. Within each cluster, DE-based mutation and crossover operations are applied to generate diverse and well-distributed synthetic samples while preserving the underlying data distribution. Additionally, a selective sampling and noise reduction mechanism is employed to filter out low-impact synthetic samples based on their contribution to classification performance. The effectiveness of ClusterDEBO is evaluated on 44 benchmark datasets using k-Nearest Neighbors (kNN), decision tree (DT), and support vector machines (SVM) as classifiers. The results demonstrate that ClusterDEBO consistently outperforms existing oversampling techniques, leading to improved class separability and enhanced classifier robustness. Moreover, statistical validation using the Friedman test confirms the significance of the improvements, ensuring that the observed gains are not due to random variations. The findings highlight the potential of cluster-assisted differential evolution as a powerful strategy for handling imbalanced datasets.
- Open Access
- 30 August 2025
- Published by International Journal of 3D Printing Technologies and Digital Industry in International Journal of 3D Printing Technologies and Digital Industry
- Vol. 9 (2) , 320-330
BODE: A BOUNDARY AWARE DIFFERENTIAL EVOLUTION BASED HYBRID OVERSAMPLING TECHNIQUE FOR IMBALANCED DATA CLASSIFICATION
Class imbalance presents a persistent challenge in supervised learning, often degrading classifier performance on underrepresented classes. This study introduces BODE, a hybrid oversampling method that combines boundary-aware instance selection, differential evolution-based perturbation, and density-constrained filtering. By targeting critical minority instances near decision boundaries, BODE generates diverse yet structurally valid synthetic samples. Experiments on 44 benchmark datasets using k-NN, Decision Tree, and SVM classifiers demonstrate that BODE consistently outperforms eleven widely used oversampling methods. Evaluated solely using the AUC metric, BODE achieves the highest average performance across all classifiers, with 28, 33, and 26 dataset-level wins, respectively. These results confirm BODE’s robustness and generalization capability, particularly in challenging scenarios involving overlapping or sparse decision regions.
- 30 August 2025
- Published by International Journal of 3D Printing Technologies and Digital Industry in International Journal of 3D Printing Technologies and Digital Industry
- Vol. 9 (2) , 320-330
Open AccessBODE: A BOUNDARY AWARE DIFFERENTIAL EVOLUTION BASED HYBRID OVERSAMPLING TECHNIQUE FOR IMBALANCED DATA CLASSIFICATION
Class imbalance presents a persistent challenge in supervised learning, often degrading classifier performance on underrepresented classes. This study introduces BODE, a hybrid oversampling method that combines boundary-aware instance selection, differential evolution-based perturbation, and density-constrained filtering. By targeting critical minority instances near decision boundaries, BODE generates diverse yet structurally valid synthetic samples. Experiments on 44 benchmark datasets using k-NN, Decision Tree, and SVM classifiers demonstrate that BODE consistently outperforms eleven widely used oversampling methods. Evaluated solely using the AUC metric, BODE achieves the highest average performance across all classifiers, with 28, 33, and 26 dataset-level wins, respectively. These results confirm BODE’s robustness and generalization capability, particularly in challenging scenarios involving overlapping or sparse decision regions.
- Open Access
- 30 June 2025
- Published by Isparta Uygulamali Bilimler Universitesi in Uluslararası Sürdürülebilir Mühendislik ve Teknoloji Dergisi
- Vol. 9 (1) , 41-50
KONTROLLÜ DENGESİZLİK SENARYOLARINDA TOPLULUK ÖĞRENME MODELLERİN SİSTEMATİK KARŞILAŞTIRMASI
Bu çalışma, sınıf dengesizliğinin topluluk öğrenme algoritmaları üzerindeki etkisini kontrollü bir deneysel tasarım ile incelemeyi amaçlamaktadır. Çalışma kapsamında, Iris ve Wine veri setleri üzerinde dört farklı sınıf dağılımı senaryosu (orijinal, hafif, orta ve şiddetli dengesizlik) uygulanmış ve her senaryoda Random Forest, Gradient Boosting ve Bagging algoritmaları test edilmiştir. Değerlendirmelerde yalnızca doğruluk değil, aynı zamanda Macro-F1, Balanced Accuracy, G-Mean ve Cohen Kappa gibi çoklu performans metrikleri kullanılmıştır. Elde edilen bulgular, Gradient Boosting modelinin yüksek dengesizlik düzeylerinde ciddi performans kayıpları yaşadığını; buna karşılık Random Forest algoritmasının tüm senaryolarda kararlı ve güvenilir sonuçlar sunduğunu ortaya koymuştur. Bu yönüyle çalışma, sınıf dengesizliğine karşı dayanıklı model seçiminin ve çok boyutlu metriklerle yapılan değerlendirmelerin önemini vurgulamaktadır.
- 30 June 2025
- Published by Isparta Uygulamali Bilimler Universitesi in Uluslararası Sürdürülebilir Mühendislik ve Teknoloji Dergisi
- Vol. 9 (1) , 41-50
Open AccessKONTROLLÜ DENGESİZLİK SENARYOLARINDA TOPLULUK ÖĞRENME MODELLERİN SİSTEMATİK KARŞILAŞTIRMASI
Bu çalışma, sınıf dengesizliğinin topluluk öğrenme algoritmaları üzerindeki etkisini kontrollü bir deneysel tasarım ile incelemeyi amaçlamaktadır. Çalışma kapsamında, Iris ve Wine veri setleri üzerinde dört farklı sınıf dağılımı senaryosu (orijinal, hafif, orta ve şiddetli dengesizlik) uygulanmış ve her senaryoda Random Forest, Gradient Boosting ve Bagging algoritmaları test edilmiştir. Değerlendirmelerde yalnızca doğruluk değil, aynı zamanda Macro-F1, Balanced Accuracy, G-Mean ve Cohen Kappa gibi çoklu performans metrikleri kullanılmıştır. Elde edilen bulgular, Gradient Boosting modelinin yüksek dengesizlik düzeylerinde ciddi performans kayıpları yaşadığını; buna karşılık Random Forest algoritmasının tüm senaryolarda kararlı ve güvenilir sonuçlar sunduğunu ortaya koymuştur. Bu yönüyle çalışma, sınıf dengesizliğine karşı dayanıklı model seçiminin ve çok boyutlu metriklerle yapılan değerlendirmelerin önemini vurgulamaktadır.
- Open Access
- 19 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 50464-50479
Optimizing Artificial Neural Networks Using Mountain Gazelle Optimizer
The performance of artificial neural networks heavily depends on the optimization of network parameters, specifically weights and biases, during the training process. Effectively adjusting these parameters is essential to minimize the error between predicted and actual outputs. While traditional training algorithms, such as gradient-based methods, have been widely used, they often face challenges like premature convergence and stagnation in local optima. Training an ANN can, therefore, be viewed as an optimization problem, where the goal is to fine-tune parameters to achieve accurate and efficient performance. In this study, we introduce a novel approach to optimizing neural network parameters using the Mountain Gazelle Optimizer (MGO), a nature-inspired metaheuristic algorithm that mimics the social hierarchy and behavioral patterns of wild mountain gazelles. The MGO algorithm leverages its unique features, including hierarchical social structure and adaptive movement strategies, to effectively navigate the complex parameter space of neural networks. The algorithm’s search mechanism integrates four key behavioral strategies: Territorial Solitary Males (TSM) for refining optimal solutions, Maternity Herds (MH) for balancing exploration and exploitation, Bachelor Male Herds (BMH) for global exploration, and Migration to Search for Food (MSF) for introducing randomness to prevent stagnation in local optima. These mechanisms work collaboratively, ensuring a dynamic and balanced optimization process throughout the training phase. To evaluate the effectiveness of the proposed approach, we conducted comprehensive experiments using various classification datasets from the UCI repository. The performance of the MGO-based neural network optimizer was compared with traditional backpropagation and several state-of-the-art optimization algorithms. Experimental results demonstrate that the proposed method exhibits superior performance in terms of convergence speed and avoiding local optima, suggesting that MGO is a promising alternative for optimizing artificial neural networks during training.
- 19 March 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 50464-50479
Open AccessOptimizing Artificial Neural Networks Using Mountain Gazelle Optimizer
The performance of artificial neural networks heavily depends on the optimization of network parameters, specifically weights and biases, during the training process. Effectively adjusting these parameters is essential to minimize the error between predicted and actual outputs. While traditional training algorithms, such as gradient-based methods, have been widely used, they often face challenges like premature convergence and stagnation in local optima. Training an ANN can, therefore, be viewed as an optimization problem, where the goal is to fine-tune parameters to achieve accurate and efficient performance. In this study, we introduce a novel approach to optimizing neural network parameters using the Mountain Gazelle Optimizer (MGO), a nature-inspired metaheuristic algorithm that mimics the social hierarchy and behavioral patterns of wild mountain gazelles. The MGO algorithm leverages its unique features, including hierarchical social structure and adaptive movement strategies, to effectively navigate the complex parameter space of neural networks. The algorithm’s search mechanism integrates four key behavioral strategies: Territorial Solitary Males (TSM) for refining optimal solutions, Maternity Herds (MH) for balancing exploration and exploitation, Bachelor Male Herds (BMH) for global exploration, and Migration to Search for Food (MSF) for introducing randomness to prevent stagnation in local optima. These mechanisms work collaboratively, ensuring a dynamic and balanced optimization process throughout the training phase. To evaluate the effectiveness of the proposed approach, we conducted comprehensive experiments using various classification datasets from the UCI repository. The performance of the MGO-based neural network optimizer was compared with traditional backpropagation and several state-of-the-art optimization algorithms. Experimental results demonstrate that the proposed method exhibits superior performance in terms of convergence speed and avoiding local optima, suggesting that MGO is a promising alternative for optimizing artificial neural networks during training.
- 15 January 2025
- Published by Ondokuz Mayis University in Black Sea Journal of Agriculture
- Vol. 8 (1) , 73-79
Farm Assistant Counts Sheep
Small livestock farming in our country is mostly based on pasture. The most important advantage of this situation is that it reduces feed expenses and increases our profitability within the farm. However, the most important problem is in the counting of animals when they come from the pasture to the pen and when they go from the pen to the pasture. This situation depends on the shepherd's attention and follow-up. However, finding experienced shepherds in our country is becoming more and more difficult every day. It may be difficult or even impossible for a sheep giving birth in the pasture to follow the herd when the geographical conditions become difficult. Quick counting of sheep and lambs as the animals enter and exit the pen depends on the shepherd's practice and experience. In order for this situation to be more realistic and to prevent personal mistakes, different alternatives should be considered. For this reason, a system has been developed using deep learning techniques to automatically count the animals in the herd when entering the pen. This system will automatically count the animals at the entrance and exit of the farm, and in case of missing animals, the system users will automatically notify the system users via web and mobile applications. With the implementation of this system, it will be possible to determine the losses that will occur on the farm with an early warning system. In our study, animals will be detected with the deep learning-based YoloV8 pre-trained model on images taken from fixed cameras that will be placed at the entrance and exit of the pen. Counting results obtained from the developed system can be used on different devices by providing multi-platform support. By disseminating this practice, losses of sheep and lambs in the pasture can be prevented.
- 15 January 2025
- Published by Ondokuz Mayis University in Black Sea Journal of Agriculture
- Vol. 8 (1) , 73-79
Farm Assistant Counts Sheep
Small livestock farming in our country is mostly based on pasture. The most important advantage of this situation is that it reduces feed expenses and increases our profitability within the farm. However, the most important problem is in the counting of animals when they come from the pasture to the pen and when they go from the pen to the pasture. This situation depends on the shepherd's attention and follow-up. However, finding experienced shepherds in our country is becoming more and more difficult every day. It may be difficult or even impossible for a sheep giving birth in the pasture to follow the herd when the geographical conditions become difficult. Quick counting of sheep and lambs as the animals enter and exit the pen depends on the shepherd's practice and experience. In order for this situation to be more realistic and to prevent personal mistakes, different alternatives should be considered. For this reason, a system has been developed using deep learning techniques to automatically count the animals in the herd when entering the pen. This system will automatically count the animals at the entrance and exit of the farm, and in case of missing animals, the system users will automatically notify the system users via web and mobile applications. With the implementation of this system, it will be possible to determine the losses that will occur on the farm with an early warning system. In our study, animals will be detected with the deep learning-based YoloV8 pre-trained model on images taken from fixed cameras that will be placed at the entrance and exit of the pen. Counting results obtained from the developed system can be used on different devices by providing multi-platform support. By disseminating this practice, losses of sheep and lambs in the pasture can be prevented.
- Open Access
- 31 December 2024
- Published by Izmir Akademi Dernegi in Journal of Metaverse
- Vol. 4 (2) , 157-164
Application of Prompt Engineering Techniques to Optimize Information Retrieval in the Metaverse
Prompt engineering techniques are instructions that enable large language models (LLMs) to solve real-world problems more effectively. These techniques enhance the capabilities of LLMs to generate accurate and efficient responses. Our study examines the challenge of acquiring comprehensive and efficient information in the metaverse through the application of various prompt engineering techniques. The main objective is to improve the accuracy and effectiveness of metaverse-related responses by leveraging LLM capabilities. In this study, 100 questions were generated using GPT, GEMINI, QWEN, and MISTRAL language models focusing on the metaverse. Our experiments indicated that responses often included unrelated information, highlighting the need for prompt engineering techniques. We applied knowledge-based, rule-based, few-shot, and template-based prompt engineering techniques to refine the responses. The performance of GPT, GEMINI, QWEN, and MISTRAL models were evaluated based on criteria including accuracy, timeliness, comprehensiveness, and consistency. Our findings reveal that prompt engineering techniques significantly enhance the efficacy of LLMs in providing improved information retrieval and response generation, aiding users in efficiently acquiring information in complex environments like the metaverse.
- 31 December 2024
- Published by Izmir Akademi Dernegi in Journal of Metaverse
- Vol. 4 (2) , 157-164
Open AccessApplication of Prompt Engineering Techniques to Optimize Information Retrieval in the Metaverse
Prompt engineering techniques are instructions that enable large language models (LLMs) to solve real-world problems more effectively. These techniques enhance the capabilities of LLMs to generate accurate and efficient responses. Our study examines the challenge of acquiring comprehensive and efficient information in the metaverse through the application of various prompt engineering techniques. The main objective is to improve the accuracy and effectiveness of metaverse-related responses by leveraging LLM capabilities. In this study, 100 questions were generated using GPT, GEMINI, QWEN, and MISTRAL language models focusing on the metaverse. Our experiments indicated that responses often included unrelated information, highlighting the need for prompt engineering techniques. We applied knowledge-based, rule-based, few-shot, and template-based prompt engineering techniques to refine the responses. The performance of GPT, GEMINI, QWEN, and MISTRAL models were evaluated based on criteria including accuracy, timeliness, comprehensiveness, and consistency. Our findings reveal that prompt engineering techniques significantly enhance the efficacy of LLMs in providing improved information retrieval and response generation, aiding users in efficiently acquiring information in complex environments like the metaverse.