Graduate School of Informatics

This study investigates the global discourse organization of Turkish TEDx talks. 70 TEDx Talks in Turkish with, reliable human-generated transcriptions were chosen to be annotated. These were collected as subtitle files and manually annotated to map out significant discourse segments. For every talk, a number of features such as the number of total words, specific transition words, duration (second-wise), speed, average embedding and the ending percentile of each sentence were used for training Machine Learning (ML) models. The results indicate that the transitions between motivational discourse segments can be predicted with an F1-score of 0.78.

Date: 22.01.2024 / 13:00 Place: B-116

Translating natural language sentences into logical expressions has been a challenging task due to contextual information and the variational complexity of sentences. In recent years, a new deep learning architecture, namely the Transformer architecture, has been providing new ways to handle what was hard or seemed impossible in natural language processing tasks. The Transformer architecture and language models that are based on it revolutionized the artificial intelligence field of research and changed how we approach natural language processing tasks. In this thesis, we conduct experiments to see whether successful results can be achieved using Transformer models in translating sentences into first-order logic expressions.

Date: 23.01.2024 / 11:00 Place: B-116

Steel production serves as the backbone of countless infrastructure projects and industrial applications worldwide. In order to maintain and improve its productivity, quality and environmental sustainability, hyperspectral imaging is a promising technology for steel industry.  A novel, non-destructive approach is presented to quantify the free lime content in steel slag by utilizing an integrated algorithm applied to hyperspectral images. This method includes spectral unmixing for mixture component quantification and endmember extraction of mixture. Methodology involved various experiments with both fresh and six-month-aged steel slag, demonstrating its accuracy compared to the Rietveld Analysis of X-ray Diffraction patterns.

Date: 11.01.2024 / 09:30 Place: B-116

State-of-the-art omics technologies use network-based contextualization methods to give molecular information about different biological contexts, like disease states, patients, and drug changes. In the beginning, this thesis identified challenging issues such as missing points in contextualization, hidden knowledge in omics datasets, bias in reference networks, and noisy interactions with highly connected nodes or hubs. Subsequently, to address these challenges, we developed pyPARAGON (PAgeRAnk-flux on Graphlet-guided network for multi-Omics data integratioN). Also, a novel tool, pyPARAGON, contextualized patient datasets by inferring patient-specific networks and complex diseases by constructing disease models, namely breast cancer and autism spectrum disorders.

Date: 22.01.2024 / 14:00 Place: A-108

In the United States, institutional investors submit 13F reports to SEC every quarter, disclosing the number of shares they own. However, investors may delay filing these reports to conceal their strategies. In this study, data from 13F filings were enriched with additional information such as stock prices and industries to create a bipartite, dynamic, and rich-attributed graph. The changes in network metrics, edge counts, and motif counts were visualized, highlighting crises in the relevant time period. The study found that clustering coefficient, significant position changes, late-filed 13F reports, and specific motif counts of graph shows significant changes during market crashes.

Date: 18.01.2024 / 10:00 Place: A-212

Cancers of unknown primary (CUP) pose treatment challenges due to unidentified primary sites, affecting survival rates significantly. Traditional methods falter in pinpointing origins. Gene expression-based studies offer promise; machine learning models trained on annotated cancer data achieve remarkable accuracy, surpassing conventional methods. Our study, utilizing three machine learning models trained on TCGA data, attained competitive accuracy, around 96%.

Date: 18.01.2024 / 11:00 Place: A-108

Despite impressive achievements made by machine learning algorithms (especially in deep learning), they are easily tricked by modified input data. Adversarial attacks target machine learning models severely. Adversarial training is an effective method against adversarial attacks, but it is not suitable for network domains due to network flow characteristics. In this thesis, the autoencoder's reconstruction error is used for detecting adversarial attacks. The IDS model, RAIDS is proposed and achieves respectful results against adversarial attacks.

Date: 17.01.2024 / 14:00 Place: A-108

Global climate change induces lake level fluctuations, impacted by evolving meteorological factors and water use. Input or output changes swiftly affect the water balance equation. This study explores predictive models for climatic and hydrologic variables, assessing their correlations with lake water level and water quality. Using diverse algorithms—Naive Method, ANN, and RNN—LSTM excels in accuracy by RMSE. Comparisons with the Naïve Method confirm ANN and RNN predictive prowess, especially with extended horizons. Correlations with temperature and evaporation highlight lake water quality impacts. Models and metrics construct a decision support tool for water managers.

Date: 19.12.2023 / 13:30 Place: A-212