Scopus İndeksli Yayınlar Koleksiyonu
Permanent URI for this collectionhttps://hdl.handle.net/20.500.12573/395
Browse
49 results
Search Results
Article Probing Adversarial Robustness of Protein Language Models: A Reproducible Case Study of ESM-2 Under Substitution-Based Attacks(AIUB Office of Research and Publication, 2026) Aydin, Zafer; Niloy, Md. Robiul Islam; Moazzam, Md. GolamArticle An Attention-Based Autoencoder Model with Gated Recurrent Unit for Stock Price Movement Prediction(Springer Nature, 2026) Kolukisa, Burak; Bakir-Gungor, Burcu; Akkas, HuseyinPredicting stock movements is crucial for investors looking to maximize their profits in rapidly changing financial markets. However, noise in stock price data makes it harder to detect the trends and ultimately decreases the performance of predictive models. To address these challenges that are caused by noise, this study proposes two novel models, i.e., an Attention-based Autoencoder (ABA) and an Attention-based Variational Autoencoder with Gated Recurrent Units (AGRUA), which uniquely integrate denoising, attention, and GRU layers. Firstly, a data set is created by adding 9 different technical indicators to the historical data of Borsa Istanbul (BIST 30) stocks. Secondly, proposed methods and other well-known deep learning models were used to remove noise from the data sets. Finally, each denoised dataset was fed separately to the Extreme Gradient Boosting model and subjected to a buy-sell process. The results were measured using trading indicators such as amount of the profits and Sharpe Ratio, Sortino Ratio and Maximum Drawdown. The proposed models produced substantial financial gains, with AGRUA achieving the highest total profit and ABA achieving the lowest average Maximum Drawdown, thereby demonstrating superior risk-adjusted performance. Lastly, Friedman and Nemenyi tests confirmed that AGRUA and ABA surpass most of the benchmarks in profit and risk-adjusted returns. The performance of the proposed method demonstrates its value in capturing nonlinear market patterns and improving decision accuracy, emphasizing the need for noise reduction before forecasting.Conference Object Citation - Scopus: 2Combining Classifiers for Protein Secondary Structure Prediction(Institute of Electrical and Electronics Engineers Inc., 2017-09) Aydin, Zafer; Uzut, Ömmu GülsümArticle GraphUnet-SS: A Novel Deep Learning Model for Protein Secondary Structure Prediction Based on U-Net Architecture(Elsevier Ltd, 2026-04) Aydin, Zafer; Görmez, Yasin; Sabzekar, MostafaArticle Frequency-Based Deep Occlusion Awareness Instance Segmentation(MDPI, 2026-02-26) Guzel, Yasin; Aydin, Zafer; Talu, Muhammed FatihOne major challenge faced by deep learning-based methods that detect target objects in the form of bounding boxes is object occlusion. High degrees of occlusion significantly diminish the accuracy of instance segmentation. Nonetheless, complex-valued Fourier descriptors can robustly represent object boundaries using minimal information. In this study, the impact of integrating Fourier descriptors-renowned for their strong representational capacity-with deep network models (UNet) that exhibit high generalization performance on instance segmentation accuracy was investigated. Within the scope of the research, nine network models were designed based on different strategies for utilizing frequency components. These variants fall into four strategy families: (i) UNet-style spectrum regression on fixed low-frequency windows (FUNet), (ii) magnitude-guided frequency selection/ROI construction (FUNet-Thr, FUNet-BBox), (iii) sequence models over tokenized FFT coefficients (BiLSTM Patch/Sorted), and (iv) encoder-only spectrum predictors with different depth/capacity (EncoderFFT1/2). To fairly evaluate the models' performance in segmenting objects subjected to disruptive factors (e.g., occlusion, blurring, noise), a specialized synthetic dataset was prepared. The task is formulated as single-target (single-instance), single-class segmentation. This dataset, automatically generated according to initial parameter values, contains images of objects moving at various speeds within a single frame. Among these models, the one termed FUNet, which relies on partial matching of central frequency components, achieved the highest segmentation accuracy despite the disruptive effects. Under the challenging Dataset 8 setting, the proposed FUNet achieved the highest overlap-based performance (Dice = 0.9329, IoU = 0.8842) among Attention U-Net, U-Net, and FourierNet, with statistically significant gains confirmed by paired per-image tests.Article BrAIn: A Comprehensive Artificial Intelligence-Based Morphology Analysis System for Brain Organoids and Neuroscience(Wiley, 2026-03-12) Polatli, Elifsu; Guner, Huseyin; Bastanlar, Yalin; Karakulah, Gokhan; Evranos, Ali Eren; Kahveci, Burak; Guven, SinanHuman-induced pluripotent stem cells (iPSCs) offer transformative potential for biomedical research, with iPSC-derived organoids providing more physiologically relevant models than traditional 2D cell cultures. Among these, brain organoids (BO) are particularly valuable for drug screening, disease modeling, and investigations into molecular pathways. Accurate representation of brain morphology is critical, as more complex organoid structures better mimic the human brain. Deep learning (DL) and machine learning (ML) approaches have become integral to analyzing organoid morphology, yet tools for comprehensive, time-resolved assessments are scarce. Here, we introduce BrAIn, a DL-based application for analyzing the developmental progression of BOs. BrAIn tracks their evolution from embryoid bodies (EBs) and quantifies parameters including area, Feret diameter, perimeter, roundness, and circularity. It also classifies budding and abnormal morphologies of 3D organoids and detects monolayer neural rosette structures, key features of neuronal differentiation. Designed with accessibility in mind, BrAIn provides a no-code interface, enabling researchers of all technical backgrounds to conduct advanced morphological analyses with ease. Our study demonstrates the application of BrAIn to evaluate the effects of different growth conditions-static, orbital shaker, and microfluidic chip-based-on BO development. Orbital shaker cultures resulted in the largest organoids, while chip-based systems achieved more homogeneous growth. Both conditions produced organoids with greater morphological complexity compared to static culture. BrAIn emerges as a robust, user-friendly tool to quantify BO development and explore how versatile growth conditions influence their morphology and maturation.Conference Object Benchmarking AI-Based Forecasting Models Across Multiple Energy Sources: A Time Series Analysis in the European Context(Institute of Electrical and Electronics Engineers Inc., 2025-09-17) Dolar, Ayça; Çinarer, GökalpArticle Non-Contact Acoustic Screening for Sleep Apnea: A Subject-Aware Deep Learning Approach(Springer Science and Business Media Deutschland GmbH, 2026-02-11) Aygün Çakıroğlu, M.; Kizilkaya Aydoǧan, E.; Bolatturk, Ö.F.; Aydoğan, S.; Ismailoǧullari, S.; Delice, Y.Purpose: To explore the feasibility of using camera-derived, non-contact audio synchronized with PSG for clinically relevant sleep-apnea classification, and to benchmark compact deep models under a subject-aware design using a previously unstudied, real-world dataset. Methods: Thirty-two adults underwent simultaneous polysomnography (PSG) and camera-based non-contact audio recording. The synchronized audio segments were used to train and compare three compact deep-learning architectures (convolutional, attention-augmented, and transformer-based) under a subject-aware evaluation design that prevented identity leakage. Model performance and calibration were assessed at both segment and subject levels using standard statistical tests. Results: Subject-level evaluation was based on a very small, imbalanced test set of six subjects (one positive). Within this limited yet previously unstudied local dataset, the CNN_trans model achieved an apparent perfect ranking performance (AUC = 1.00; 95% CI 0.00–1.00), though this likely reflects the small, imbalanced test cohort, with recall = 1.00 and precision = 0.55. The wide confidence interval reflects substantial statistical uncertainty, and DeLong comparisons showed no significant AUC difference between CNN_trans and CNN_att (ΔAUC = − 0.042; p = 0.43). Conclusion: PSG-synchronized, non-contact audio supports accurate and well-calibrated sleep-apnea classification with compact deep models. This subject-aware evaluation suggests that contactless acoustic monitoring may have potential clinical relevance, motivating larger, multi-site validation. © The Author(s), under exclusive licence to Springer Nature Switzerland AG 2026.Conference Object Citation - Scopus: 1Words Speak Louder Than Actions: Decoding Emotions Through NLP(Institute of Electrical and Electronics Engineers Inc., 2024-10-26) Paksoy, Melda; Bakal, GokhanEmotion detection in text remains a significant challenge in Natural Language Processing due to human emotions' complexity and subtle nuances. This paper presents multiple experimental models for emotion classification using an up-to-date dataset curated to address 13 emotions implied in Twitter posts. We evaluated various machine learning (ML) models, including Logistic Regression, Random Forest, SVM, and XGBoost, alongside deep learning (DL) architectures such as LSTM and CNN. Our results demonstrate the efficacy of deep learning models, particularly the CNN model by achieving an impressive F1 score of 0.99. This study contributes to emotion detection capabilities, paving the way for more nuanced and accurate sentiment analysis (SA) in various text analysis applications. © 2025 Elsevier B.V., All rights reserved.Conference Object Text Classification Experiments on Contextual Graphs Built by N-Gram Series(Springer International Publishing AG, 2025) Sen, Tarik Uveys; Yakit, Mehmet Can; Gumus, Mehmet Semih; Abar, Orhan; Bakal, GokhanTraditional n-gram textual features, commonly employed in conventional machine learning models, offer lower performance rates on high-volume datasets compared to modern deep learning algorithms, which have been intensively studied for the past decade. The main reason for this performance disparity is that deep learning approaches handle textual data through the word vector space representation by catching the contextually hidden information in a better way. Nonetheless, the potential of the n-gram feature set to reflect the context is open to further investigation. In this sense, creating graphs using discriminative ngram series with high classification power has never been fully exploited by researchers. Hence, the main goal of this study is to contribute to the classification power by including the long-range neighborhood relationships for each word in the word embedding representations. To achieve this goal, we transformed the textual data by employing n-gram series into a graph structure and then trained a graph convolution network model. Consequently, we obtained contextually enriched word embeddings and observed F1-score performance improvements from 0.78 to 0.80 when we integrated those convolution-based word embeddings into an LSTM model. This research contributes to improving classification capabilities by leveraging graph structures derived from discriminative n-gram series.