Mustafa Gokce Baydogan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Mustafa Gokce Baydogan has a total of 51 co-authors. Here are the top co-authors by number of shared publications:
George Runger
School of Computing, Informatics and Decision Systems Engineering, Arizona State University, Tempe, USA
3 shared publicationsJing Li
Arizona State University
3 shared publications- Berk Görgülü
DeGroote School of Business, McMaster University, Hamilton, Ontario L8S 4M4, Canada
2 shared publications - Emel Şeyma Küçükaşcı
Department of Industrial Engineering, Istanbul Commerce University, Istanbul, Turkey
2 shared publications - Erhan Can Ozcan
Boston University
2 shared publications - Hadi Jahanshahi
Data Science Lab Ryerson University M5B 1G3 Toronto ON Canada
2 shared publications
Latest Publications
- Open Access
- 29 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 309-325
Gradient-Boosted Decision Tree Optimizer for Antenna Optimization
The use machine learning-assisted optimization methods in the design of antennas have been increasing. Although neural networks (NNs) and Gaussian process regression (GPR) are widely used, their scalability to higher dimensions poses several challenges, such as the requirement for excessive data, extensive hyper-parameter tuning, and longer training times. In contrast, gradient-boosted decision trees (GBDTs) exhibit superior performance with limited training data, along with faster training and more efficient hyper-parameter tuning. Motivated by these advantages, we introduce a GBDT-assisted optimization (GBDTO) algorithm tailored for high dimensional problems. Beginning with an initial sample set, GBDTO builds a GBDT model and sequentially samples the input parameter space while searching for an optimal objective value. Compared to Bayesian optimization (BO) and NN-assisted optimization (ONN), GBDTO achieves faster convergence and superior objective values, as demonstrated through benchmarks using the Black-Box Optimization Benchmarking test suite, and several antenna designs. Numerical experiments across 480 instances of 12-dimensional 24 functions demonstrate 13% and 31% improvement in mean rank count over BO and ONN, respectively. Moreover, high dimensional antenna design examples indicate more than 50% faster convergence for a given optimization target and 7 − 54% improvement in the objective value for a fixed number of iterations, compared to BO and ONN. In addition to its optimization efficiency, GBDTO offers inherent and efficient feature importance analysis, which is extremely useful for user guidance.
- 29 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 309-325
Open AccessGradient-Boosted Decision Tree Optimizer for Antenna Optimization
The use machine learning-assisted optimization methods in the design of antennas have been increasing. Although neural networks (NNs) and Gaussian process regression (GPR) are widely used, their scalability to higher dimensions poses several challenges, such as the requirement for excessive data, extensive hyper-parameter tuning, and longer training times. In contrast, gradient-boosted decision trees (GBDTs) exhibit superior performance with limited training data, along with faster training and more efficient hyper-parameter tuning. Motivated by these advantages, we introduce a GBDT-assisted optimization (GBDTO) algorithm tailored for high dimensional problems. Beginning with an initial sample set, GBDTO builds a GBDT model and sequentially samples the input parameter space while searching for an optimal objective value. Compared to Bayesian optimization (BO) and NN-assisted optimization (ONN), GBDTO achieves faster convergence and superior objective values, as demonstrated through benchmarks using the Black-Box Optimization Benchmarking test suite, and several antenna designs. Numerical experiments across 480 instances of 12-dimensional 24 functions demonstrate 13% and 31% improvement in mean rank count over BO and ONN, respectively. Moreover, high dimensional antenna design examples indicate more than 50% faster convergence for a given optimization target and 7 − 54% improvement in the objective value for a fixed number of iterations, compared to BO and ONN. In addition to its optimization efficiency, GBDTO offers inherent and efficient feature importance analysis, which is extremely useful for user guidance.
- Open Access
Smooth Ranking SVM via Cutting-Plane Method
The most popular classification algorithms are designed to maximize classification accuracy during training. However, this strategy may fail in the presence of class imbalance since it is possible to train models with high accuracy by overfitting to the majority class. On the other hand, the Area Under the Curve (AUC) is a widely used metric to compare classification performance of different algorithms when there is a class imbalance, and various approaches focusing on the direct optimization of this metric during training have been proposed. Among them, SVM-based formulations are especially popular as this formulation allows incorporating different regularization strategies easily. In this work, we develop a prototype learning approach that relies on cutting-plane method, similar to Ranking SVM, to maximize AUC. Our algorithm learns simpler models by iteratively introducing cutting planes, thus overfitting is prevented in an unconventional way. Furthermore, it penalizes the changes in the weights at each iteration to avoid large jumps that might be observed in the test performance, thus facilitating a smooth learning process. Based on the experiments conducted on 73 binary classification datasets, our method yields the best test AUC in 25 datasets among its relevant competitors.
Open AccessSmooth Ranking SVM via Cutting-Plane Method
The most popular classification algorithms are designed to maximize classification accuracy during training. However, this strategy may fail in the presence of class imbalance since it is possible to train models with high accuracy by overfitting to the majority class. On the other hand, the Area Under the Curve (AUC) is a widely used metric to compare classification performance of different algorithms when there is a class imbalance, and various approaches focusing on the direct optimization of this metric during training have been proposed. Among them, SVM-based formulations are especially popular as this formulation allows incorporating different regularization strategies easily. In this work, we develop a prototype learning approach that relies on cutting-plane method, similar to Ranking SVM, to maximize AUC. Our algorithm learns simpler models by iteratively introducing cutting planes, thus overfitting is prevented in an unconventional way. Furthermore, it penalizes the changes in the weights at each iteration to avoid large jumps that might be observed in the test performance, thus facilitating a smooth learning process. Based on the experiments conducted on 73 binary classification datasets, our method yields the best test AUC in 25 datasets among its relevant competitors.
- 16 November 2023
- Published by Elsevier in European Journal of Operational Research
- Vol. 314 (1) , 297-307
Column generation-based prototype learning for optimizing area under the receiver operating characteristic curve
- 16 November 2023
- Published by Elsevier in European Journal of Operational Research
- Vol. 314 (1) , 297-307
Column generation-based prototype learning for optimizing area under the receiver operating characteristic curve
- Open Access
nTreeClus: a Tree-based Sequence Encoder for Clustering Categorical Series
The overwhelming presence of categorical/sequential data in diverse domains emphasizes the importance of sequence mining. The challenging nature of sequences proves the need for continuing research to find a more accurate and faster approach providing a better understanding of their (dis)similarities. This paper proposes a new Model-based approach for clustering sequence data, namely nTreeClus. The proposed method deploys Tree-based Learners, k-mers, and autoregressive models for categorical time series, culminating with a novel numerical representation of the categorical sequences. Adopting this new representation, we cluster sequences, considering the inherent patterns in categorical time series. Accordingly, the model showed robustness to its parameter. Under different simulated scenarios, nTreeClus improved the baseline methods for various internal and external cluster validation metrics for up to 10.7% and 2.7%, respectively. The empirical evaluation using synthetic and real datasets, protein sequences, and categorical time series showed that nTreeClus is competitive or superior to most state-of-the-art algorithms.
Open AccessnTreeClus: a Tree-based Sequence Encoder for Clustering Categorical Series
The overwhelming presence of categorical/sequential data in diverse domains emphasizes the importance of sequence mining. The challenging nature of sequences proves the need for continuing research to find a more accurate and faster approach providing a better understanding of their (dis)similarities. This paper proposes a new Model-based approach for clustering sequence data, namely nTreeClus. The proposed method deploys Tree-based Learners, k-mers, and autoregressive models for categorical time series, culminating with a novel numerical representation of the categorical sequences. Adopting this new representation, we cluster sequences, considering the inherent patterns in categorical time series. Accordingly, the model showed robustness to its parameter. Under different simulated scenarios, nTreeClus improved the baseline methods for various internal and external cluster validation metrics for up to 10.7% and 2.7%, respectively. The empirical evaluation using synthetic and real datasets, protein sequences, and categorical time series showed that nTreeClus is competitive or superior to most state-of-the-art algorithms.
- 22 April 2022
- Published by Elsevier in Neurocomputing
- Vol. 494, 224-241
nTreeClus: A tree-based sequence encoder for clustering categorical series
- 22 April 2022
- Published by Elsevier in Neurocomputing
- Vol. 494, 224-241
nTreeClus: A tree-based sequence encoder for clustering categorical series
- 11 January 2022
- Published by Springer Nature in Journal of Global Optimization
- Vol. 83 (4) , 639-670
Multiple instance classification via quadratic programming
- 11 January 2022
- Published by Springer Nature in Journal of Global Optimization
- Vol. 83 (4) , 639-670
Multiple instance classification via quadratic programming
- 1 December 2021
- Published by Elsevier in Pattern Recognition
- Vol. 120
Randomized trees for time series representation and similarity
- 1 December 2021
- Published by Elsevier in Pattern Recognition
- Vol. 120
Randomized trees for time series representation and similarity
- 23 October 2021
- Published by Elsevier in Pattern Recognition
- Vol. 123, 108393
Semi-supervised extensions of multi-task tree ensembles
- 23 October 2021
- Published by Elsevier in Pattern Recognition
- Vol. 123, 108393
Semi-supervised extensions of multi-task tree ensembles
- Open Access
Learning Maximally Predictive Prototypes in Multiple Instance Learning
In this work, we propose a simple model that provides permutation invariant maximally predictive prototype generator from a given dataset, which leads to interpretability of the solution and concrete insights to the nature and the solution of a problem. Our aim is to find out prototypes in the feature space to map the collection of instances (i.e. bags) to a distance feature space and simultaneously learn a linear classifier for multiple instance learning (MIL). Our experiments on classical MIL benchmark datasets demonstrate that proposed framework is an accurate and efficient classifier compared to the existing approaches.
Open AccessLearning Maximally Predictive Prototypes in Multiple Instance Learning
In this work, we propose a simple model that provides permutation invariant maximally predictive prototype generator from a given dataset, which leads to interpretability of the solution and concrete insights to the nature and the solution of a problem. Our aim is to find out prototypes in the feature space to map the collection of instances (i.e. bags) to a distance feature space and simultaneously learn a linear classifier for multiple instance learning (MIL). Our experiments on classical MIL benchmark datasets demonstrate that proposed framework is an accurate and efficient classifier compared to the existing approaches.
- 22 April 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 20 (12) , 4635-4642
Estimation of Influence Distance of Bus Stops Using Bus GPS Data and Bus Stop Properties
In congested cities where commuting time doubles during peak hours, it is crucial to identify every network problem. Here, it is aimed to analyze the traffic behavior of the buses around the bus stops and to show the effects of the surrounding network to speedup the pattern of buses using the trajectory data of Global Positioning System (GPS)-equipped bus fleet. As an indicator of speed differences near bus stops, influence distance is suggested, wherein the speed of buses significantly changes while approaching and leaving the bus stops. Speed patterns of buses operating on 12 bus routes in Istanbul are analyzed. The data include more than 5000 daily trajectory log files and 25 million rows of location and time information during April 2016. The influence distance, measured using the fused lasso method, for the 438 bus stops varies from 36 to 174m, with an average value of 98m. In the second part of this paper, correlation of the influence distances of the bus stops with the surrounding interruptions is investigated. A distance matrix of surrounding interruptions to the nearest bus stop is generated. This matrix is used as parameters of M5 Prime, random forest, and extremely randomized trees models in order to predict the influence distances. The models show that the passenger demand plays a key role in the influence distances of the bus stops. It is also found that the changes in the number of lanes and the location of the traffic lights are quite effective on the influence distances.
- 22 April 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Intelligent Transportation Systems
- Vol. 20 (12) , 4635-4642
Estimation of Influence Distance of Bus Stops Using Bus GPS Data and Bus Stop Properties
In congested cities where commuting time doubles during peak hours, it is crucial to identify every network problem. Here, it is aimed to analyze the traffic behavior of the buses around the bus stops and to show the effects of the surrounding network to speedup the pattern of buses using the trajectory data of Global Positioning System (GPS)-equipped bus fleet. As an indicator of speed differences near bus stops, influence distance is suggested, wherein the speed of buses significantly changes while approaching and leaving the bus stops. Speed patterns of buses operating on 12 bus routes in Istanbul are analyzed. The data include more than 5000 daily trajectory log files and 25 million rows of location and time information during April 2016. The influence distance, measured using the fused lasso method, for the 438 bus stops varies from 36 to 174m, with an average value of 98m. In the second part of this paper, correlation of the influence distances of the bus stops with the surrounding interruptions is investigated. A distance matrix of surrounding interruptions to the nearest bus stop is generated. This matrix is used as parameters of M5 Prime, random forest, and extremely randomized trees models in order to predict the influence distances. The models show that the passenger demand plays a key role in the influence distances of the bus stops. It is also found that the changes in the number of lanes and the location of the traffic lights are quite effective on the influence distances.