Ömür Bali
Ömür Bali has a total of 8 co-authors. Here are the top co-authors by number of shared publications:
- Hacer Turgut
ILab
4 shared publications - Afra Arslan
ILab
3 shared publications - Arda Yücel
ILab
3 shared publications - Gülfem Işıklar Alptekin
Galatasaray University
2 shared publications - Günce Keziban Orman
Galatasaray University
2 shared publications - Tan Doruk Yetki
ILab
2 shared publications
Latest Publications
- Open Access
- 19 June 2026
- Published by Sakarya University Journal of Computer and Information Sciences in Sakarya University Journal of Computer and Information Sciences
- No. Advanced O,p. 690-699
Feature Extraction for Real Estate Images and Titles with LLMs
Images and titles often contain rich latent information about their associated objects, particularly on web-based platforms. Real estate websites provide a clear example, where listing images and titles provide important details that assist users in their decision-making. However, these unstructured elements cannot be directly utilized in downstream machine learning tasks, since their contextual meaning is not directly interpretable. This work aims to transform listing images and titles into structured, tabular representations, making them suitable for analytical and predictive modeling. To this end, we propose a modular framework based on state-of-the-art large language models. The framework incorporates ReAct, LLM-as-a-Judge, and few-shot prompting techniques. Its performance is evaluated on a real-world real estate dataset and compared with BERT and CLIP-based baselines. Experimental results demonstrate that our framework achieves up to a 44.26% improvement in recall for listing attributes, such as the presence of a balcony or the furnishing status of a property.
- 19 June 2026
- Published by Sakarya University Journal of Computer and Information Sciences in Sakarya University Journal of Computer and Information Sciences
- No. Advanced O,p. 690-699
Open AccessFeature Extraction for Real Estate Images and Titles with LLMs
Images and titles often contain rich latent information about their associated objects, particularly on web-based platforms. Real estate websites provide a clear example, where listing images and titles provide important details that assist users in their decision-making. However, these unstructured elements cannot be directly utilized in downstream machine learning tasks, since their contextual meaning is not directly interpretable. This work aims to transform listing images and titles into structured, tabular representations, making them suitable for analytical and predictive modeling. To this end, we propose a modular framework based on state-of-the-art large language models. The framework incorporates ReAct, LLM-as-a-Judge, and few-shot prompting techniques. Its performance is evaluated on a real-world real estate dataset and compared with BERT and CLIP-based baselines. Experimental results demonstrate that our framework achieves up to a 44.26% improvement in recall for listing attributes, such as the presence of a balcony or the furnishing status of a property.
- 24 January 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 310
Beyond clicks: Measuring attractiveness and satisfaction in e-commerce using Bayesian models with conversion signals
- 24 January 2026
- Published by Elsevier in Expert Systems with Applications
- Vol. 310
Beyond clicks: Measuring attractiveness and satisfaction in e-commerce using Bayesian models with conversion signals
- Open Access
A Comprehensive Framework for Measuring the Immediate Impact of TV Advertisements: TV-Impact
Measuring the immediate impact of television advertisements (TV ads) on online traffic poses significant challenges in many aspects. Nonetheless, a comprehensive consideration is essential to fully grasp consumer reactions to TV ads. So far, the measurement of this effect has not been studied to a large extent. Existing studies have either determined how a specific focus group, i.e., toddlers, people of a certain age group, etc., react to ads via simple statistical tests using a case study approach or have examined the effects of advertising with simple regression models. This study introduces a comprehensive framework called TV-Impact. The framework uses a Bayesian structural time-series model called CausalImpact. There are additional novel approaches developed within the framework. One of the novelties of TV-Impact lies in its dynamic algorithm for selecting control variables which are supporting data sources and presumed to be unaffected by TV ads. In addition, we proposed the concept of Group Ads to combine overlapping ads into a single ad structure. Then, Random Forest Regressor, which is a commonly preferred supervised learning method, is used to decompose the impact into single ads. The TV-Impact framework was applied to the data of iLab, a venture company in Turkey, and manages its companies’ advertising strategies. The findings reveal that the TV-Impact model positively influenced the companies’ strategies for allocating their TV advertisement budgets and increased the amount of traffic driven to company websites, serving as an effective decision support system.
Open AccessA Comprehensive Framework for Measuring the Immediate Impact of TV Advertisements: TV-Impact
Measuring the immediate impact of television advertisements (TV ads) on online traffic poses significant challenges in many aspects. Nonetheless, a comprehensive consideration is essential to fully grasp consumer reactions to TV ads. So far, the measurement of this effect has not been studied to a large extent. Existing studies have either determined how a specific focus group, i.e., toddlers, people of a certain age group, etc., react to ads via simple statistical tests using a case study approach or have examined the effects of advertising with simple regression models. This study introduces a comprehensive framework called TV-Impact. The framework uses a Bayesian structural time-series model called CausalImpact. There are additional novel approaches developed within the framework. One of the novelties of TV-Impact lies in its dynamic algorithm for selecting control variables which are supporting data sources and presumed to be unaffected by TV ads. In addition, we proposed the concept of Group Ads to combine overlapping ads into a single ad structure. Then, Random Forest Regressor, which is a commonly preferred supervised learning method, is used to decompose the impact into single ads. The TV-Impact framework was applied to the data of iLab, a venture company in Turkey, and manages its companies’ advertising strategies. The findings reveal that the TV-Impact model positively influenced the companies’ strategies for allocating their TV advertisement budgets and increased the amount of traffic driven to company websites, serving as an effective decision support system.
- Open Access
- 21 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 5253-5254
Prod2Vec-Var: A Session Based Recommendation System with Enhanced Diversity
Understanding user behavior and leveraging this information in recommendation systems pose challenges for websites lacking a login system or with limited logged-in users. This study introduces the Prod2Vec-Var recommendation system, a modified version of a session-based recommendation algorithm aimed at enhancing the performance of product recommendation systems with a cold-start extension. The proposed model builds upon the original prod2vec algorithm, incorporating an additional step to improve the diversity of product recommendations. The project entails a well-designed data pipeline, effectively processing user actions to align them with the model, and implementing unique functions that expand the range of products capturing users' attention. Moreover, a straightforward yet effective cold-start model is developed to address newly added products that have not been viewed by users. The outcome of our project, namely product suggestions, is presented to users of cimri.com, one of iLab's affiliated companies, which attracts millions of daily visits, thereby enabling seamless access to desired products. Experimental results demonstrated the superior performance of our model compared to the other two different strategies in running recommendation on popular products, as evidenced by favorable R@1, R@5, R@10, and R@15 metrics. Concerning less popular products, we observed an improvement in our model's performance as the value of K increased, ultimately achieving optimal results in terms of R@15. Additionally, our cold-start model for new products substantiated the efficacy of our methodology, yielding the highest scores across R@5, R@10, and R@15 metrics.
- 21 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 5253-5254
Open AccessProd2Vec-Var: A Session Based Recommendation System with Enhanced Diversity
Understanding user behavior and leveraging this information in recommendation systems pose challenges for websites lacking a login system or with limited logged-in users. This study introduces the Prod2Vec-Var recommendation system, a modified version of a session-based recommendation algorithm aimed at enhancing the performance of product recommendation systems with a cold-start extension. The proposed model builds upon the original prod2vec algorithm, incorporating an additional step to improve the diversity of product recommendations. The project entails a well-designed data pipeline, effectively processing user actions to align them with the model, and implementing unique functions that expand the range of products capturing users' attention. Moreover, a straightforward yet effective cold-start model is developed to address newly added products that have not been viewed by users. The outcome of our project, namely product suggestions, is presented to users of cimri.com, one of iLab's affiliated companies, which attracts millions of daily visits, thereby enabling seamless access to desired products. Experimental results demonstrated the superior performance of our model compared to the other two different strategies in running recommendation on popular products, as evidenced by favorable R@1, R@5, R@10, and R@15 metrics. Concerning less popular products, we observed an improvement in our model's performance as the value of K increased, ultimately achieving optimal results in terms of R@15. Additionally, our cold-start model for new products substantiated the efficacy of our methodology, yielding the highest scores across R@5, R@10, and R@15 metrics.