Cemal Okan Sakar
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Cemal Okan Sakar has a total of 49 co-authors. Here are the top co-authors by number of shared publications:
- Olcay Kursun
Department of Computer Science and Computer Information Systems, Auburn University at Montgomery, Montgomery, AL 36117, USA
20 shared publications - Gorkem Serbes
Department of Biomedical Engineering, Yildiz Technical University, Istanbul 34220, Türkiye
8 shared publications Fikret Gurgen
7 shared publications- Nizamettin Aydin
Department of Computer Engineering, Istanbul Technical University, Istanbul, Turkey
6 shared publications Betul Erdogdu Sakar
5 shared publications- Fikret Gurgen
Department of Computer Engineering, Boğaziçi University, İstanbul, Turkey
3 shared publications
Latest Publications
- 21 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-8
Corporate Banking Product Purchase Prediction Across Multiple Horizons Using Multimodal Tabular Features
Corporate banking product recommendation studies generally focus on single-step prediction and often rely on raw sequential transaction data as model input. In this paper, we examine an alternative tabular prediction framework with three main characteristics. First, we extend the prediction setting from the immediate next period to 1-, 2-, and 3-month horizons through cumulative horizon definitions. Second, we build the prediction models on aggregated multimodal client features rather than raw event sequences. Third, we analyze the resulting tabular models with SHAP in order to provide an interpretable view of the main drivers behind product purchase predictions. We construct the feature space from transaction records, geolocation events, and dialog embeddings obtained from the publicly available Multimodal Banking Dataset (MBD-mini). We further enrich this feature space with lag-based and cumulative temporal variables to represent client-level behavioral history. Our experimental design includes modality ablation (96 experiments), model comparison (48 experiments), and 5-fold cross-validation (60 experiments). We obtain the best results with XGBoost using the full multimodal feature set. This model achieves a mean ROC-AUC of 0.792 for the 1 -month horizon and exceeds the aggregation baseline value of 0.783 reported in the original benchmark study. Our SHAP analysis indicates that cumulative activity measures, transaction accumulation patterns, and recent client behavior are among the most influential predictors. We also find that the proposed framework remains competitive across extended horizons. The maximum performance decrease is limited to $\mathbf{2. 7 8}$ percentage points at the $\mathbf{3}$-month horizon. These findings suggest that aggregated multimodal features can provide an effective and interpretable basis for lightweight multi-horizon prediction in corporate banking.- 21 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 1-8
Corporate Banking Product Purchase Prediction Across Multiple Horizons Using Multimodal Tabular Features
Corporate banking product recommendation studies generally focus on single-step prediction and often rely on raw sequential transaction data as model input. In this paper, we examine an alternative tabular prediction framework with three main characteristics. First, we extend the prediction setting from the immediate next period to 1-, 2-, and 3-month horizons through cumulative horizon definitions. Second, we build the prediction models on aggregated multimodal client features rather than raw event sequences. Third, we analyze the resulting tabular models with SHAP in order to provide an interpretable view of the main drivers behind product purchase predictions. We construct the feature space from transaction records, geolocation events, and dialog embeddings obtained from the publicly available Multimodal Banking Dataset (MBD-mini). We further enrich this feature space with lag-based and cumulative temporal variables to represent client-level behavioral history. Our experimental design includes modality ablation (96 experiments), model comparison (48 experiments), and 5-fold cross-validation (60 experiments). We obtain the best results with XGBoost using the full multimodal feature set. This model achieves a mean ROC-AUC of 0.792 for the 1 -month horizon and exceeds the aggregation baseline value of 0.783 reported in the original benchmark study. Our SHAP analysis indicates that cumulative activity measures, transaction accumulation patterns, and recent client behavior are among the most influential predictors. We also find that the proposed framework remains competitive across extended horizons. The maximum performance decrease is limited to $\mathbf{2. 7 8}$ percentage points at the $\mathbf{3}$-month horizon. These findings suggest that aggregated multimodal features can provide an effective and interpretable basis for lightweight multi-horizon prediction in corporate banking.- Open Access
- 9 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 22211-22231
A Systematic Approach to Key Phrase Extraction for Turkish: Adapting Embedding-Based Models
Key phrase extraction (KPE) is the identification of informative phrases to provide a summary of a given document. Although unsupervised methods are highly valued for their domain independence and generalization capabilities, research in this area has predominantly focused on English texts. Conversely, KPE for agglutinative languages like Turkish remains understudied, primarily due to the challenges posed by complex morphology and a scarcity of high-quality, organized benchmarks. This study explores a systematic approach to adapting existing embedding-based KPE models, which have proven successful in English, for the Turkish language. A heuristic-based extraction method is investigated for constructing candidate pools, with attention to the impact of chunking strategies. The analysis includes three embedding-based models (AttentionRank, SAMRank, and JointGL) evaluated on two newly compiled Turkish datasets from the Information and Health Sciences domains. We also examine how well these architectures adapt and discuss their relevance for future research. While statistical methods remain strong baselines, the experimental results show that Turkish pre-trained language models (PLMs) enable existing embedding-based models to be adapted to Turkish and support the establishment of new baselines. Among these approaches, AttentionRank shows the highest performance and stability. In addition, hybrid methods that combine embedding-based and statistical techniques achieve strong performance, adapt well to different dataset properties, and provide robust baselines for future studies.
- 9 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 22211-22231
Open AccessA Systematic Approach to Key Phrase Extraction for Turkish: Adapting Embedding-Based Models
Key phrase extraction (KPE) is the identification of informative phrases to provide a summary of a given document. Although unsupervised methods are highly valued for their domain independence and generalization capabilities, research in this area has predominantly focused on English texts. Conversely, KPE for agglutinative languages like Turkish remains understudied, primarily due to the challenges posed by complex morphology and a scarcity of high-quality, organized benchmarks. This study explores a systematic approach to adapting existing embedding-based KPE models, which have proven successful in English, for the Turkish language. A heuristic-based extraction method is investigated for constructing candidate pools, with attention to the impact of chunking strategies. The analysis includes three embedding-based models (AttentionRank, SAMRank, and JointGL) evaluated on two newly compiled Turkish datasets from the Information and Health Sciences domains. We also examine how well these architectures adapt and discuss their relevance for future research. While statistical methods remain strong baselines, the experimental results show that Turkish pre-trained language models (PLMs) enable existing embedding-based models to be adapted to Turkish and support the establishment of new baselines. Among these approaches, AttentionRank shows the highest performance and stability. In addition, hybrid methods that combine embedding-based and statistical techniques achieve strong performance, adapt well to different dataset properties, and provide robust baselines for future studies.
- Open Access
- 24 January 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Asian and Low-Resource Language Information Processing
Seq2Turk: Turkish Spelling Error Correction Using Context-Dependent Sequence-to-Sequence Model
Spelling errors in digital content negatively affect the understanding of the readers and also the effectiveness of Natural Language Processing (NLP) tasks applied on such data. In this study, we propose a context-dependent spelling correction model for Turkish, which is an agglutinative language with a rich morphology and a complex grammatical structure. For this purpose, we design a sequence-to-sequence architecture based on a pre-trained RoBERTa model. In our end-to-end framework, we first generated a clean dataset using a highly accurate spelling error detection model. Then, we developed functions to introduce artificial spelling errors based on common mistakes in digital media and trained a transformer-based encoder-decoder model to correct the errors in this dataset. The experimental analyses were conducted on a real dataset consisting of user-generated content obtained from digital platforms. The quantitative and qualitative results in comparison with existing spelling correction models proposed for Turkish and recent generative chatbots, including chatGPT and Google Gemini, show the effectiveness of the proposed model.
- 24 January 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Asian and Low-Resource Language Information Processing
Open AccessSeq2Turk: Turkish Spelling Error Correction Using Context-Dependent Sequence-to-Sequence Model
Spelling errors in digital content negatively affect the understanding of the readers and also the effectiveness of Natural Language Processing (NLP) tasks applied on such data. In this study, we propose a context-dependent spelling correction model for Turkish, which is an agglutinative language with a rich morphology and a complex grammatical structure. For this purpose, we design a sequence-to-sequence architecture based on a pre-trained RoBERTa model. In our end-to-end framework, we first generated a clean dataset using a highly accurate spelling error detection model. Then, we developed functions to introduce artificial spelling errors based on common mistakes in digital media and trained a transformer-based encoder-decoder model to correct the errors in this dataset. The experimental analyses were conducted on a real dataset consisting of user-generated content obtained from digital platforms. The quantitative and qualitative results in comparison with existing spelling correction models proposed for Turkish and recent generative chatbots, including chatGPT and Google Gemini, show the effectiveness of the proposed model.
- 19 October 2025
- Published by Springer Nature
- p. 1255-1274
Transaction Count Prediction in Correspondent Banking Using a Hybrid CNN-LSTM Neural Network
- 19 October 2025
- Published by Springer Nature
- p. 1255-1274
Transaction Count Prediction in Correspondent Banking Using a Hybrid CNN-LSTM Neural Network
- 14 October 2025
- Published by Springer Nature
- p. 69-83
Making Data AI-Ready: A Comprehensive Framework for Automated Data Pipelines
- 14 October 2025
- Published by Springer Nature
- p. 69-83
Making Data AI-Ready: A Comprehensive Framework for Automated Data Pipelines
- Open Access
- 28 February 2025
- Published by Springer Nature in Applied Intelligence
- Vol. 55 (7) , 1-18
A statistical categorization-based curriculum learning approach for multi-task classification of images
Image classification and the detection of features within images remain significant challenges in computer vision. Several approaches, including serial task models and multi-output models, have been explored to address these challenges. This study focuses on multitasking attention mechanisms, which enable simultaneous categorization of data and tasks. By applying a statistical framework, the proposed method enhances the efficiency and accuracy of image classification and feature detection, with a focus on handling multiple tasks concurrently. To enhance the robustness of the model, a data-driven approach based on curriculum learning was proposed. The experiments were conducted using two distinct datasets. The first dataset involves forensic examinations, specifically identifying firearms and their calibers from firing pin marks. The proposed model achieved an accuracy of 95% in brand detection and 98% in caliber detection on this dataset. In the second part of the experiments, the animals with attributes 2 (AwA2) dataset, where state-of-the-art models have previously been applied, was used. The proposed model reduced classification errors by 1 to 10% compared to traditional convolutional neural network (CNN) architectures. The experimental results from both the forensic and public datasets demonstrate that the proposed model effectively handles multitask classification tasks, validating its applicability across diverse domains.
- 28 February 2025
- Published by Springer Nature in Applied Intelligence
- Vol. 55 (7) , 1-18
Open AccessA statistical categorization-based curriculum learning approach for multi-task classification of images
Image classification and the detection of features within images remain significant challenges in computer vision. Several approaches, including serial task models and multi-output models, have been explored to address these challenges. This study focuses on multitasking attention mechanisms, which enable simultaneous categorization of data and tasks. By applying a statistical framework, the proposed method enhances the efficiency and accuracy of image classification and feature detection, with a focus on handling multiple tasks concurrently. To enhance the robustness of the model, a data-driven approach based on curriculum learning was proposed. The experiments were conducted using two distinct datasets. The first dataset involves forensic examinations, specifically identifying firearms and their calibers from firing pin marks. The proposed model achieved an accuracy of 95% in brand detection and 98% in caliber detection on this dataset. In the second part of the experiments, the animals with attributes 2 (AwA2) dataset, where state-of-the-art models have previously been applied, was used. The proposed model reduced classification errors by 1 to 10% compared to traditional convolutional neural network (CNN) architectures. The experimental results from both the forensic and public datasets demonstrate that the proposed model effectively handles multitask classification tasks, validating its applicability across diverse domains.
- 3 May 2023
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 82 (28) , 44259-44275
Concealed pistol detection from thermal images with deep neural networks
- 3 May 2023
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 82 (28) , 44259-44275
Concealed pistol detection from thermal images with deep neural networks
- 5 March 2023
- Published by Elsevier in Knowledge-Based Systems
- Vol. 266
Creating CREATE queries with multi-task deep neural networks
- 5 March 2023
- Published by Elsevier in Knowledge-Based Systems
- Vol. 266
Creating CREATE queries with multi-task deep neural networks
- Open Access
- 30 January 2021
- Published by Marmara University in International Journal of Advances in Engineering and Pure Sciences
- Vol. 33 (1) , 18-27
Stock Market Prediction by Combining Stock Price Information and Sentiment Analysis
Predicting the stock market instrument price is a valuable but challenging machine learning task. Researchers use advanced techniques to improve the generalization ability of stock prediction models. However, considering that the stock market highly depends on the political and macroeconomic developments as well as the mood of the related investors, the models that use only stock prices fail to cover all factors affecting the stock market. Therefore, to improve the prediction accuracy of stock market prediction, in this study, we first apply sentiment analysis to the news related with the market and related stock, and then combine the sentiment labels of the news with stock prices and commonly used technical indicators. The obtained cumulative dataset is used to train a long short-term memory recurrent neural network, and the output of this regression model is used in the prediction of the closing price movement to decide whether the closing price next day will be higher. The experiments performed on 8-year data showed that while the f1 score of the model built without sentiment analysis was around 0.56, it has increased to 0.65 when stock prices are combined with sentiment labels. The results show that the model with sentiment labels fits better to the actual prices especially when there is a high volatility in the stock price.
- 30 January 2021
- Published by Marmara University in International Journal of Advances in Engineering and Pure Sciences
- Vol. 33 (1) , 18-27
Open AccessStock Market Prediction by Combining Stock Price Information and Sentiment Analysis
Predicting the stock market instrument price is a valuable but challenging machine learning task. Researchers use advanced techniques to improve the generalization ability of stock prediction models. However, considering that the stock market highly depends on the political and macroeconomic developments as well as the mood of the related investors, the models that use only stock prices fail to cover all factors affecting the stock market. Therefore, to improve the prediction accuracy of stock market prediction, in this study, we first apply sentiment analysis to the news related with the market and related stock, and then combine the sentiment labels of the news with stock prices and commonly used technical indicators. The obtained cumulative dataset is used to train a long short-term memory recurrent neural network, and the output of this regression model is used in the prediction of the closing price movement to decide whether the closing price next day will be higher. The experiments performed on 8-year data showed that while the f1 score of the model built without sentiment analysis was around 0.56, it has increased to 0.65 when stock prices are combined with sentiment labels. The results show that the model with sentiment labels fits better to the actual prices especially when there is a high volatility in the stock price.
- 1 November 2020
- Published by Elsevier in Applied Soft Computing
- Vol. 96
Distracted driver detection by combining in-vehicle and image data using deep learning
- 1 November 2020
- Published by Elsevier in Applied Soft Computing
- Vol. 96
Distracted driver detection by combining in-vehicle and image data using deep learning