Theses supervised by Doç. Dr. Oğuz Ata
10 theses · Altınbaş University
Kantitatif hisse senedi analiz uygulaması
Bu tez çalışmasında, Kamuyu Aydınlatma Platformu (KAP) ve Türkiye Bankalar Birliği (TBB) gibi resmî kaynaklardan elde edilen çoklu finansal verilerin, ASP.NET Core tabanlı katmanlı bir mimari üzerinden Playwright ve Selenium ile zamanlanmış veri çekimi, otomatik ön işleme, SQL veri tabanı kayıt altyapısı ve Python destekli görselleştirme bileşenleri aracılığıyla entegre edildiği uçtan uca otomatik bir kantitatif analiz sistemi geliştirilmiştir. Sistem hem reel sektör hem de finans sektöründeki kurumlara yönelik olmak üzere Kantitatif, Banka, Sigorta ve Faktöring başlıklı dört farklı PDF raporu üretmekte; ayrıca günlük piyasa özetlerini oluşturmakta ve Telegram bot entegrasyonu aracılığıyla gerçek zamanlı bildirimler sağlamaktadır. Elde edilen deneysel bulgular, çok kaynaklı veri entegrasyonunun analitik kapsamı anlamlı ölçüde genişlettiğini; Excel ve PDF modüllerinin ise raporların detay seviyesini önemli ölçüde artırdığını göstermektedir. Performans ölçümleri, sistemin raporlama süresini %65'e kadar azalttığını ve sektör bazlı rating karşılaştırmaları sayesinde portföy kararlarında duyarlılığı ve doğruluğu artırdığını ortaya koymuştur. Tartışma bölümünde veri güncellemelerinde senkronizasyon sorunları, zaman damgası uyumsuzlukları ve versiyon kontrolü gibi çoklu kaynak yönetimi zorlukları analiz edilmiştir. Sistem mimarisi, API tabanlı servisler, gömülebilir bileşenler ve çoklu kiracı yapısı sayesinde, ticarileştirilmeye uygun tam ölçekli bir SaaS (Software as a Service) ürünü niteliği taşımaktadır. Sonuç olarak geliştirilen sistem; veri doğruluğu, işlem hızı ve kullanıcı etkileşimini tek bir platform altında birleştirerek, Türkiye finansal piyasaları için ölçeklenebilir, sürdürülebilir ve genişletilebilir bir karar destek aracı sunmaktadır.
Kardiyovasküler hastalık tahmininin geliştirilmesi için entegratif makine öğrenmesi yaklaşımları: XGBoost ve ANFIS algoritmalarının karşılaştırmalı analizi
Cardiovascular diseases (CVDs) are the leading cause of death globally, underscoring the need for advanced detection and diagnostic methods to enhance patient outcomes. This study investigates the efficacy of two machine learning algorithms, XGBoost and the Adaptive Neuro-Fuzzy Inference System (ANFIS), in predicting heart disease across diverse datasets. Utilizing datasets from the UCI Machine Learning Repository, including Switzerland, Cleveland, Hungarian, Long Beach VA, and Statlog Heart, standard preprocessing techniques such as imputation, standardization, one-hot encoding, and SMOTEENN were applied to ensure consistent modeling conditions. Both models underwent extensive training and optimization. XGBoost excelled, particularly achieving 100% accuracy in the Switzerland and Statlog datasets, while ANFIS demonstrated its strength in modeling complex patterns, notably achieving perfect accuracy in the Cleveland dataset. Performance evaluations using accuracy, precision, recall, F1 score, F2 score, and ROC-AUC score highlighted XGBoost's consistent high precision and recall, vital for reliable CVD diagnosis. In contrast, ANFIS showed potential in clinical settings with its high F2 scores, emphasizing the reduction of false negatives. The study highlights the advantages of using advanced machine learning models like XGBoost and ANFIS in cardiovascular diagnostics, suggesting further research with larger and more varied datasets to refine these models and advance medical diagnostics using machine learning.
Çok etikeli dengesizlik verileri ile litolojiyi sınıflandırmak için derin öğrenme yöntemlerinin kullanılması
To anticipate and detect lithologies in a variety of surveys, geologists can save operational expenses and uptime by utilising deep learning methodologies and applications. Accurate data processing and scientific research using data gathered in different geological areas are made possible by this. The four lithologies data in the present research were analysed and classified using multi-class imbalance issues and high dimensionality. One of the biggest issues facing modern data analysis is the imbalance in data classification. Particularly when combined with other challenging factors like the existence of overlapping class distributions, and data imbalance can have a significant impact on the accuracy of classification. When there are several classes involved, mutual imbalance relationships between them exacerbate the situation, making its influence more evident. Furthermore, the high dimensionality issue may result in overfitting and increased computational complexity, both of which may impair classification efficiency. Recursive Feature Elimination (RFE) is used to find the most valuable predictive features, while Synthetic Minority Oversampling (SMOTE) is used to resample the data. Using hybrid multi-class DL system unbalanced learning approach is our solution to solving these issues. Finally, by offering precise categorization and quick responses about the interpretation of data collected in many study regions, we think that our innovations might contribute to the advancement of geological research.
Yüz tanıma sistemi için VGG tabanlı özellik çıkarma
•Facial recognition technologies are one of the main aspects of many things for example; security, biometrics, and social media. That is where we go ahead to present a feature extraction for our face recognition system based on the VGG approach. We assemble a collection of facial images and then process them to keep all the images consistent and properly set to avoid poor-quality images. The prioritized model exemplifies the use of VGG16, employed to extract high-level features from faces, that follow identification by the classification algorithm. System efficiency is evaluated concerning indicators of quality, for instance, accuracy precision, recall, and F1-Score. The results show that our model, based on feature extraction using VGG, has high accuracy and an accuracy rate with an LR model is 91%, ANN 0.87, SVM 0.89, KNN 0.74, DT0,39, GB0.75, and RF 0.74for FR. The results show that our proposed works well and is efficient in facial recognition functions. We believe that this kind of research takes facial recognition technology to a new level of development and will be a great example for other studies.
Derin öğrenme modeli ile zatürre tespiti
Zatürre, her bireyin yaşamının herhangi bir döneminde maruz kalabileceği bir akciğer enfeksiyonudur. Tedavi edilmediği takdirde ciddi sağlık sorunlarına ve hatta ölüme neden olabilir. Bu nedenle, hastalığın erken teşhisi büyük bir öneme sahiptir. Zatürrenin tespitinde çeşitli yöntemler kullanılmakla birlikte, en yaygın yöntemlerden biri akciğer röntgen görüntülerinin incelenmesidir. Bu görüntüler uzmanlar tarafından titizlikle değerlendirilir. Ancak bu sürecin doğruluğu, süresi ve harcanan emek önemli faktörler arasında yer almaktadır. Görüntü analizi alanında, yapay zekanın hızla gelişmesiyle birlikte önemli çalışmalar yapılmıştır. Bu tezde, akciğer röntgen görüntüleri kullanılarak makine öğrenmesi ile zatürre teşhisini destekleyen bir yöntem geliştirilecektir. Böylece, teşhis sürecinde zaman yönetimi, doğruluk oranı ve harcanan çabanın optimize edilmesi sağlanacaktır. Literatürdeki çalışmalar ve geliştirilen modeller dikkate alınarak, evrişimsel sinir ağı tabanlı yöntemler incelenecektir. Çalışmada kullanılan akciğer görüntüleri üç gruba ayrılmıştır: eğitim, doğrulama ve test. Sonuç olarak, evrişimsel sinir ağı modellerinin kullanılmasıyla elde edilen sonuçların, aynı konu üzerinde araştırma yapan araştırmacıların sonuçları kıyaslanarak literatüre katkıda bulunması ve doktorlara karar desteği sağlaması amaçlanmaktadır.
Şahin balığı algoritması: Çift uygunluk fonksiyonlu yeni bir matematiksel optimizasyon algoritması
Inspired by the gender shift behaviour shown in Hawkfish, this thesis contributes a new concept in optimization algorithm modelling by using this natural phenomenon as a basis for solving optimization problems through mathematical modelling and imitation. To evaluate the performance of HFOA, it is initially subjected to benchmarking against various test problems, including the CEC-06 (2019) benchmark suite, to quantitatively and qualitatively validate its efficacy. Subsequently, HFOA is deployed to identify the optimal configurations for the welded beam design problem and the tension/compression spring (TCS) design problem, showcasing its practical utility. The outcomes reveal that the proposed HFOA algorithm consistently outperforms well-established and recent algorithms documented in the literature. Furthermore, the results obtained in real-world applications underscore the effectiveness of HFOA in tackling problems characterized by unknown search spaces.
Akademik bütünlükün sağlanması: insan tarafından yazılmış ve yapay zekâ tarafından üretilmiş metinleri ayırt etmek için transformer tabanlı ve toplu makine öğrenme yaklaşımları
The explosive development of large language models (LLMs) such as ChatGPT and Bard transformed text creation, offering new possibilities for academic communication and raising deep concerns regarding authenticity and authorship. As AI-created content has come to be almost indistinguishable from human writing, the need for open and reliable mechanisms to authenticate authorship has become crucial in sustaining academic trust. This work presents and illustrates two mutually complementary Natural Language Processing (NLP) frameworks the Transformer-Based Detection Framework and the Ensemble Learning-Based Detection Framework and contrasts them on three evenly balanced datasets: AI-GA, HWAI, and HAGT-1M. The former uses Sentence-BERT (SBERT) and RoBERTa embeddings together with Logistic Regression (LR) and Feed-Forward Neural Networks (FNNs) to detect deep semantic and syntactic trends. Across datasets, transformer configurations achieved strong performance: SBERT+LR: 93.66% (AI-GA), 91.37% (HWAI), 99.64% (HAGT-1M); RoBERTa+FNN: 91.13% (AI-GA), 90.29% (HWAI), 99.95% (HAGT-1M) with RoBERTa-FNN peaking at 99.95% on HAGT-1M. To ensure transparency and surmount the "black-box" limitation of transformers, the second method leverages ensemble learning by combining LR, SVM, Random Forest, Extremely Randomized Trees, XGBoost, AdaBoost, and SGD with soft voting over TF-IDF and linguistic feature sets (n-grams, lexical diversity, sentence length). The ensemble achieved 99.88% (AI-GA), 99.37% (HWAI), and 100% (HAGT-1M), outperforming prior state-of-the-art and demonstrating increased robustness on varied academic text. The current study focuses on reproducible preprocessing, data exploration, and careful threshold tuning regarding sensitivity and fairness, especially for non-native research scholars. Model linguistic authenticity as a three-dimensional property present in syntax, semantics, and stylistic coherence. Offer open, scalable, and transparent detection protocols suitable for universities, publishers, and research integrity offices. Future research will include the investigation of hybrid quantum machine learning (HQML), explainable-AI tools with more detail (e.g., SHAP/LIME/attention visualization), and multilingual extension in order to maintain cross-domain generalization.
Improving prediction of chest infections using machine learning algorithms from X-ray images
This thesis aims to present and develop a deep learning-based computer application that helps related professionals quickly and accurately identify common lung diseases by interpreting chest X-ray images. Convolutional neural network (CNN) is a very efficient technique for image processing. The proposed work has used AlexNet, and this architecture was fine-tuned and used to extract features from CXR images and analyse these features using energy spectral density individually. The four datasets from open source chest X-ray (CXR) images are publicly available, 15153, 6432, 9208, and 312, respectively, of X-ray images. The experimental results in our proposed system using SVM classifier were accurate at 95.6%, 95.4%, 97.2%, and 95.7%, respectively, for these datasets. The method was compared with several studies.
Using of convolutional neural network forgrayscale image colorization
The advancement of data technology had motivated more facilitation of human life including security and pattern recognition. Large information can be obtained from the image after proper feature extraction. Thus, the field of image processing has gained extended attention especially after incorporating artificial intelligence (AI). One of the vital problems that image processing is looking after is dealing with ancient images or in other words, the grayscale images which were captured before the introduction of RBG technology. Analyzing grayscale images for more curtail knowledge extraction is possible through colorizing technology where the grayscale image is converted into an RBG image. In this thesis, the development of an accurate image colorization approach is proposed using a convolutional neural network (CNN). Optimization of colorization performance is conducted through tuning of the CNN model. Results have shown that the proposed classifier have been scored of 80.91 Percent accuracy.
Augmented random search applied in artificial intelligence
Random search algorithms are helpful for several ill-structured world improvement issues with continuous or distinct variables. Usually, random search algorithms give perfect optimality for locating an honest resolution quickly with convergence. Random search algorithms embrace simulated annealing, tabu search, genetic algorithms, biological process programming, particle swarm improvement, pismire colony improvement, cross-entropy, random approximation, multi- begin and bunch algorithms, to call some. They will be classified as world versus native search, or instance-based versus model-based. However, one feature these ways share is that the use of likelihood in determinative their repetitious procedures. This text provides a summary of those random search algorithms, with a probabilistic read that ties them along. Augmented Random Search is actually one among the foremost mind-blowing algorithms where the fundamental is using a systematic approach particularly Augmented Random Search. It is a newly born methodology for Reinforcement Learning which supposes to use the strategy for limited contrasts and it can do precisely the same that Google Deep Mind did in their achievement a year ago, in other words, an AI to walk and keep running over a field. With the ongoing selection of standard benchmark suites, a huge assemblage of late research has connected RL strategies for constant control within recreation situations.