Welcome to the Polen - ITU Academic Archive

Polen stores all academic resources—such as books, articles, theses, papers, reports, and research data—directly or indirectly published by the academic and administrative staff and students of Istanbul Technical University in a digital environment in accordance with international standards. It facilitates monitoring the university’s academic performance, ensures the long-term preservation of resources, and provides open access to publications in compliance with copyright regulations to enhance their impact.

The ITU Open Access System provides access to scientific and artistic works—such as books, articles, theses, encyclopedias, and artworks—produced by our faculty members and students in accordance with international standards and intellectual property rights.

Supported by @SelenSoft Yazılım

Articles

Books

Book Parts

Master Theses

Doctoral Theses

Communities in DSpace

Select a community to browse its collections.

Now showing 1 - 5 of 6

Açık Ders Kaynakları 30
İTÜ'de verilen bazı derslerin video kayıtlarını içermektedir.
Akademik Çalışmalar 45561
İTÜ'de yapılan akademik çalışmaları içermektedir.
Arşiv Koleksiyonu 603
Çeşitli birimlerde yer alan arşiv koleksiyonunu içermektedir.
Sunumlar 42
Kurum içinde yapılan çeşitli sunumların çıktılarını içermektedir.
Temel Kayıtlar 2825
Master Data

Recent Submissions

Access status: Open Access ,
Kredi kartı dolandırıcılık tespitinde optuna tabanlı hiperparametre optimizasyonu ile derin öğrenme modellerinin karşılaştırılması
(Lisansüstü Eğitim Enstitüsü, 2026-06-29) Mammadli, Elvin; Özcan, Tuncay; 403241015; İşletme
Finansal sistemlerin dijitalleşmesi ve elektronik ödeme yöntemlerinin yaygınlaşmasıyla birlikte, kredi kartı dolandırıcılığı bireysel kullanıcılar, finans kurumları ve düzenleyici otoriteler için ciddi bir risk unsuru hâline gelmiştir. Geleneksel kural tabanlı sistemlerin değişen dolandırıcılık örüntüleri karşısında yetersiz kalması, veriden öğrenme yeteneğine sahip yapay zekâ tabanlı çözümleri zorunlu kılmıştır. Bu tez çalışmasında, Kaggle üzerinde yayımlanan ve yaygın olarak kullanılan Avrupa Kredi Kartı İşlemleri veri seti üzerinde, makine öğrenmesi ve derin öğrenme tabanlı dolandırıcılık tespit modelleri geliştirilmiş ve karşılaştırılmıştır. Çalışma kapsamında ilk olarak K-En yakın komşu (KNN), Naive Bayes ve Lojistik regresyon gibi klasik makine öğrenmesi yöntemleri test edilmiş; bu yöntemlerin yüksek dengesizlik altında Kesinlik değerlerinin %6–45 aralığında kaldığı ve pratik olarak yetersiz olduğu gösterilmiştir. Ardından üç farklı derin öğrenme mimarisi ele alınmıştır: tek boyutlu Evrişimli Sinir Ağı (1D-CNN), Uzun-Kısa Süreli Bellek (LSTM) ve CNN-LSTM hibrit modeli. Veri setinin %0,17 oranında dolandırıcılık örneği içeren ileri derecede dengesiz yapısı, model eğitiminde önemli bir zorluk oluşturmuştur. Bu probleme yönelik olarak SMOTE algoritmasıyla sentetik veri üretimi, sınıf ağırlıklandırma ({0:1, 1:10}) ve karar eşiği optimizasyonu gibi tamamlayıcı stratejiler bütünleşik bir çerçeve içinde uygulanmıştır. Modellerin performansını artırmak amacıyla Optuna çerçevesi kullanılarak Bayes tabanlı hiperparametre optimizasyonu gerçekleştirilmiştir. Her model için 50 deneme (trial) yürütülmüş ve hedef metrik olarak Kesinlik belirlenmiştir. Optimizasyon sürecinde filtre sayısı, LSTM birim sayısı, seyreltme oranı, öğrenme hızı ve yığın boyutu gibi kritik parametreler için en uygun kombinasyonlar tespit edilmiştir. eşik ayarlama aşamasında karar eşiği 0,50 yerine, üç modelde de toplam maliyeti minimize eden 0,98 değerine çekilmiş, böylece yanlış alarm sayısı önemli ölçüde azaltılmıştır. Deneysel bulgular, Optuna tabanlı optimizasyonun özellikle LSTM ve CNN-LSTM modellerinde belirgin iyileşmeler sağladığını ortaya koymaktadır. LSTM modelinin Kesinlik değeri %44'ten %60'a, CNN-LSTM hibrit modelinin Kesinlik değeri ise %55'ten %76'ya yükselmiştir. Varsayılan parametrelerle dahi %69 Kesinlik değerine ulaşan CNN modelinde ise optimizasyon, performansı benzer seviyede korumuştur. F1-skoru değerlerinde de tutarlı iyileşmeler gözlemlenmiş; LSTM modelinde F1- skoru değeri %58'den %70'e, CNN-LSTM hibrit modelinde ise %68'den %80'e yükselmiştir. AUC-ROC değerleri üç modelde de 0,94 üzerinde gerçekleşmiştir. xxii Üç model karşılaştırıldığında, en yüksek Kesinlik (%76), en yüksek F1-skoru (%80), en yüksek PR-AUC (0,8506) ve en düşük tahmini maliyet (8.800 €) değerlerini CNN-LSTM hibrit modeli sağlamıştır. CNN modeli en yüksek AUC-ROC değerini (0,9575) sunarken, LSTM modeli zamansal bağımlılıkları yakalamada tamamlayıcı bir rol oynamıştır. Bulgular, dolandırıcılık tespiti gibi sınıf dengesizliğinin yoğun olduğu problemlerde, hiperparametre optimizasyonu, sınıf ağırlıklandırma ve karar eşiği ayarlamasının birlikte uygulanmasının; yalnızca sentetik örnekleme veya tek başına model seçiminden çok daha etkili sonuçlar verdiğini göstermektedir. Çalışma kapsamında ayrıca, yapısal (tablo hâlindeki) verilerde güçlü referans yöntemler olarak kabul edilen iki ağaç tabanlı topluluk modeli olan Rastgele Orman ve XGBoost, aynı veri ve ön işleme koşulları altında Optuna ile optimize edilerek değerlendirilmiştir. XGBoost %96 Kesinlik ve %85 F1-skoru ile tüm modeller arasında en yüksek kesinliğe ulaşmış, yalnızca 3 yanlış alarm üretmiştir; Rastgele Orman ise 0,9821 ile en yüksek AUC-ROC değerini sağlamıştır. Bununla birlikte, her iki topluluk modeli de derin öğrenme modellerine kıyasla daha fazla dolandırıcılık vakası kaçırmış (sırasıyla 23 ve 20 yanlış negatif); benimsenen maliyet çerçevesinde en düşük toplam maliyeti (8.800 €) CNN-LSTM hibrit modeli korumuştur (XGBoost: 11.650 €, Rastgele Orman: 10.800 €). Bu bulgular, ağaç tabanlı toplulukların yapısal verilerdeki gücünü teyit ederken, kesinlik odaklı topluluklar ile duyarlılık odaklı derin mimariler arasındaki tercihin kurumun maliyet yapısına bağlı olduğunu göstermektedir. Sonuç olarak, bu tez; klasik makine öğrenmesi algoritmalarının ileri düzey sınıf dengesizliği altındaki yetersizliğini deneysel olarak ortaya koyarak derin öğrenme yaklaşımlarının üstünlüğünü göstermiş; Optuna ile yürütülen sistematik hiperparametre optimizasyonunun bu modellerin performansını önemli ölçüde artırdığını kanıtlamıştır. Önerilen yaklaşım, bankacılık sistemlerinde uygulanabilir bir çerçeve sunarak hem akademik literatüre hem de finans sektörüne pratik katkı sağlamaktadır.
Access status: Open Access ,
The effect of the EU GDPR on national advertising expenditure: A difference-in-differences analysis
(Graduate School, 2026-06-29) Sert, Koralp; Chakravorty, Shourjo; 412221013; Economics
This thesis examines whether the European Union's General Data Protection Regulation (GDPR) made advertising more expensive by protecting consumers' data. The research question is motivated by the shift of advertising toward digital, data-driven personalisation and by the GDPR's strict consent and enforcement framework, which took effect on 25 May 2018. Using national digital advertising expenditure from the IAB Europe AdEx Benchmark for 2016–2020, deflated to real 2016 euros using country-specific consumer price indices, the study implements a difference-in-differences design that compares EU member states to a set of non-EU European control countries before and after the policy. The empirical strategy relies on two-way fixed effects in a log-linear specification of real advertising spend, with progressively richer macroeconomic and sectoral controls. The baseline model estimates a statistically insignificant decline of approximately 3.9% in real EU advertising expenditure relative to the control group. Extended specifications attenuate the point estimate to about 3.0% while confidence intervals continue to include zero under conventional OLS and heteroskedasticity-robust inference. Three additional robustness checks refine this baseline narrative. First, a placebo exercise that assigns a fake treatment date in 2017 using only pre-2018 data detects a statistically significant pre-trend, indicating that part of the post-GDPR gap between EU and non-EU countries predates the regulation. Second, dropping Norway and Switzerland from the control group—jurisdictions with GDPR-equivalent or EEA-linked data rules—roughly doubles the estimated effect to about −9%, marginally significant at the 10% level, consistent with spillover-related attenuation of the baseline estimate. Third, continuous treatment-intensity specifications interacting the post period with pre-GDPR market size show significant heterogeneity across countries, rejecting a one-size-fits-all treatment effect. Overall, the evidence supports the view that GDPR strengthened privacy protections without imposing a large, uniformly detectable macroeconomic burden on aggregate digital advertising, while highlighting sensitivity to control-group composition and pre-trends. The thesis contributes the first country-panel DiD analysis of IAB Europe national advertising data with a full set of macro controls over the 2016–2020 window and discusses implications for policy evaluation in related jurisdictions. Methodologically, the analysis documents heteroskedasticity in several specifications and therefore reports heteroskedasticity-robust standard errors alongside the main tables in the Appendix. A placebo test and alternative control-group compositions complement the baseline difference-in-differences estimates. Together, these diagnostics illustrate both the strengths and the limitations of aggregate DiD evidence in a short panel where global shocks such as the COVID-19 pandemic overlap the post-treatment window. The thesis closes by outlining directions for future work, including longer post periods, richer non-European control countries, and sector-specific heterogeneity.
Access status: Open Access ,
Empathy as a non-ideal necessity: On the normative role of empathy for justice under pluralism
(Graduate School, 2026-06-30) Bitlisli, Mustafa Nurkan; Koçan, Gürcan; 419221005; Political Studies
Empathy is a process whereby affective responses congruent with another's situation are elicited through perspective-taking or direct encounter with another. Therefore, it is manifested in various milieus of everyday life where people engage with one another. The necessarily social and affective character of empathy brings forth a vast set of possibilities for political theory where the concept of justice is not strictly demarcated from emotions or care. Having been inspired by such possibilities, this study strives to investigate the normative role of empathy for advancing justice in pluralistic democracies. By combining political theory with moral philosophy and interpretative qualitative analysis, this study adopts an interdisciplinary approach for addressing how empathy is, and ought to be, manifested in political life, as well as for indicating both the potential benefits and dangers such an association between empathy and justice can bring forth. The main argument of the thesis is that empathy has a limited but indispensable role for justice in that it is an epistemic and motivational necessity for justice under non-ideal conditions wherein different forms of domination, motivational deficits and misrecognition occur. To this end, the thesis begins with a conceptual analysis of justice and empathy, mapping their terrain in order to set a framework in which the main approaches, dimensions and elements of the said concepts are clarified. Only then does this study put forward the epistemic part of its argument by claiming that without empathy, lived reasons and socially situated experiences remain systematically invisible under idealized abstractions. That which makes empathy a non-ideal necessity here is that the legitimacy of justice principles under pluralism requires the incorporation of the standpoints of all affected. Since empathy offers a non-substitutable epistemic access for disclosing social suffering and lived experiences, which are otherwise invisible, it becomes a non-ideal necessity for justice, through which the legitimacy of principles can be sustained. Subsequently, the motivational part of the main argument, which comprises two distinct yet converging lines, claims that without the motivational basis empathy provides, it is not possible to sustain solidarity and civic compliance under strain, and to extend recognition-based struggles beyond the directly harmed. Considering empathy as a motivational non-ideal necessity, at this point, is related to the premise that an inclusive form of justice is not only distributive but also relational, and it requires motivational compliance for its own stability. Empathy, by eliciting prosocial motivation through empathic concern, becomes a non-ideal necessity for sustaining the motivational compliance and recognitional struggles for justice under non-ideal conditions. However, this study does not take the contributions of empathy for granted, and emphasizes that the normative role of empathy is limited, since empathy itself can be socially mediated in ways that reproduce the forms of social suffering, and is liable to bias. Since empathy requires critical scrutiny and structural awareness, the main argument of the thesis is underpinned by a constructive proposal in which the negative understanding of justice as the elimination of injustice is promoted through the juxtaposition of empathic engagement and generalizable rules. Consequently, this study situates empathy as a catalyst for surfacing hidden harms and sustaining solidarity and recognitional struggles against exclusionary patterns, contributing to the challenges raised against liberal frameworks of justice in which material conditions and social positions are abstracted away by idealized forms of rationality, and in which emotions and affective processes are treated as distortions.
Access status: Open Access ,
Empathy as a non-ideal necessity: On the normative role of empathy for justice under pluralism
(Graduate School, 2026-06-30) Bitlisli, Mustafa Nurkan; Koçan, Gürcan; 419221005; Political Studies
Empathy is a process whereby affective responses congruent with another's situation are elicited through perspective-taking or direct encounter with another. Therefore, it is manifested in various milieus of everyday life where people engage with one another. The necessarily social and affective character of empathy brings forth a vast set of possibilities for political theory where the concept of justice is not strictly demarcated from emotions or care. Having been inspired by such possibilities, this study strives to investigate the normative role of empathy for advancing justice in pluralistic democracies. By combining political theory with moral philosophy and interpretative qualitative analysis, this study adopts an interdisciplinary approach for addressing how empathy is, and ought to be, manifested in political life, as well as for indicating both the potential benefits and dangers such an association between empathy and justice can bring forth. The main argument of the thesis is that empathy has a limited but indispensable role for justice in that it is an epistemic and motivational necessity for justice under non-ideal conditions wherein different forms of domination, motivational deficits and misrecognition occur. To this end, the thesis begins with a conceptual analysis of justice and empathy, mapping their terrain in order to set a framework in which the main approaches, dimensions and elements of the said concepts are clarified. Only then does this study put forward the epistemic part of its argument by claiming that without empathy, lived reasons and socially situated experiences remain systematically invisible under idealized abstractions. That which makes empathy a non-ideal necessity here is that the legitimacy of justice principles under pluralism requires the incorporation of the standpoints of all affected. Since empathy offers a non-substitutable epistemic access for disclosing social suffering and lived experiences, which are otherwise invisible, it becomes a non-ideal necessity for justice, through which the legitimacy of principles can be sustained. Subsequently, the motivational part of the main argument, which comprises two distinct yet converging lines, claims that without the motivational basis empathy provides, it is not possible to sustain solidarity and civic compliance under strain, and to extend recognition-based struggles beyond the directly harmed. Considering empathy as a motivational non-ideal necessity, at this point, is related to the premise that an inclusive form of justice is not only distributive but also relational, and it requires motivational compliance for its own stability. Empathy, by eliciting prosocial motivation through empathic concern, becomes a non-ideal necessity for sustaining the motivational compliance and recognitional struggles for justice under non-ideal conditions. However, this study does not take the contributions of empathy for granted, and emphasizes that the normative role of empathy is limited, since empathy itself can be socially mediated in ways that reproduce the forms of social suffering, and is liable to bias. Since empathy requires critical scrutiny and structural awareness, the main argument of the thesis is underpinned by a constructive proposal in which the negative understanding of justice as the elimination of injustice is promoted through the juxtaposition of empathic engagement and generalizable rules. Consequently, this study situates empathy as a catalyst for surfacing hidden harms and sustaining solidarity and recognitional struggles against exclusionary patterns, contributing to the challenges raised against liberal frameworks of justice in which material conditions and social positions are abstracted away by idealized forms of rationality, and in which emotions and affective processes are treated as distortions.
Access status: Metadata only ,
The preparation of ferroboron and ferrovanadium by Aluminothermic reduction
(1996) Yücel O.; O., Yücel