Graduate School of Informatics

This study proposes a model to explain dominant word orders emergence and distributions. It integrates preferences observed in newly emerged small deaf communities and explores how new behaviours are adopted through iterated learning. The model examines the transmission of word order preferences through generations, simulates reproduction and community development, and considers possible innate biases. It analyses the effects of language pressures, and network structures on generations. Various scenarios with different parameters are presented, including community size, bias distribution, communication networks, and pressure effects. Results are provided to explain the outcomes.

Date: 21.07.2023 / 13:00 Place: A-212

In this study, the performance of machine learning methods in credit risk scoring for "Buy Now Pay Later" customers is compared with the performance of standard credit risk models. Both traditional credit risk models and machine learning algorithms are evaluated using a real dataset. The comparison of models is conducted through variable selection, model training, and performance metrics. The results summarize to what extent machine learning methods outperform traditional models in credit risk assessment for "Buy Now Pay Later" customers. It is expected that this study will provide practical recommendations to improve risk assessment processes for financial institutions and credit providers.

Date: 14.07.2023 / 13:30 Place: A-212

This thesis compares the performance of univariate and multivariate energy consumption forecasts using deep learning techniques. The study finds that the univariate model outperforms the multivariate models for two of the three data sets tested. Among all the model architectures, LSTM outperforms all the univariate experiments, while TFT performs best among the multivariate experiments. The results suggest that univariate models are superior in forecasting energy consumption despite being less complex and requiring significantly less training time, cost, and resources.

Date: 13.08.2023 / 13:15 Place: A-212

Understanding biological phenomena is not always easy. Some phenomena cannot be easily transferred due to their complexity. Various visualization methods are used to speed up the process. These include new technologies such as Virtual and Mixed Reality. Such new technologies offer new approaches in the field of education by giving users completely new interactive experiences. In this study, we interactively simulated the defense of white blood cells, one of the body's defense mechanisms, in a 3D environment. We developed the application on 3 different platforms and obtained data about the platforms that users would prefer for such an application.

Date: 18.07.2023 / 10:30 Place: A-212

This thesis introduces a novel loss function, the Class Distance Weighted Cross Entropy (CDW-CE) loss, for automated severity assessment of Ulcerative colitis (UC) using Convolutional Neural Networks (CNN) on endoscopic images. CDW-CE considers ordinal relationships between classes and enhances prediction accuracy, outperforming other loss functions across different metrics and architectures. It also improves class activation maps' precision, aiding explanation of model predictions. The approach's broad applicability is confirmed by successful testing on a diabetic retinopathy dataset. The study also created the largest public UC image dataset.

Date: 17.07.2023 / 12:30 Place: A-212

The study introduces a new method to classify Hepatocellular Carcinoma (HCC) using a hyperspectral imaging system (HSI) combined with a light microscope. This method leverages 3D convolutions in Convolutional Neural Networks (CNNs) to train a robust classifier, capturing unique spectral and spatial features automatically. The approach also addresses class imbalance in the dataset by employing a focal loss function, preventing overfitting. The results show that hyperspectral data surpasses RGB data in liver cancer tissue classification, and enhanced spectral resolution improves accuracy, highlighting the importance of both spectral and spatial features for effective cancer tissue classification.

Date: 19.06.2023 / 15:45 Place: B-116

This thesis proposes automatic evaluation techniques to assess mobile health applications as alternatives to manual methods. It utilizes large language models to classify applications' employed Persuasive System Design (PSD) principles from collected user reviews and application descriptions. Mobile Application Rating Scale (MARS) scores are predicted with regression models trained on classification probabilities that are enriched with additional descriptive data. The study’s proposed techniques outperform baseline models, while feature importance scores show that PSD principles have significant contributions to quality predictions of mobile applications.

Date: 20.06.2023 / 14:00 Place: A-212

This study aims to gain an understanding of the content and the epistemic causal connective acquisition process of children aged 6;5- 8 in Turkish. For this purpose, to test whether there are connectives that children use specific to domains of causality as adults do (Çokal, Zeyrek, & Sanders, 2020); child and adult participants are given both descriptive (biased for content relations) and argumentative (biased for epistemic relations) tasks. Data that is collected from these tasks are annotated and statistically analyzed.

Date: 27.04.2023 / 15:30 Place: A-212