4
Archived Theses
0
DOIs Assigned
0%
DOI Rate
Archived Theses
Learning from rare clicks: A comparative study of imbalance-handling techniques for deep CTR prediction
Click-through Rate (CTR) prediction is an essential but difficult task in online advertising and recommendation systems. The imbalanced nature of click data, where positive click instances are much fewer in number than non-click instances, is one of the major issues in this task. This work presents an extensive comparison of rare event handling methods combined with three state-of-the-art deep learning models for CTR prediction tasks, namely, Deep and Cross Network (DCN), Deep Factorization Machine (DeepFM), and Deep Neural Network (DNN). The rare event handling methods being considered are data-level methods such as Synthetic Minority Over-sampling Technique (SMOTE) and NearMiss undersampling, and algorithm-level methods such as class weight, PolyLoss, and Dice loss functions. The experiments are conducted on two large real-world datasets: the Avazu CTR prediction dataset, containing 1.4 million training samples and a CTR of 16.16 percent; and the Trendyol JFY Ads dataset, a 5 million training sample dataset provided by the company Trendyol, an e-commerce company, where the CTR is around 5 percent. The experimental results, validated through multi-seed Wilcoxon signed-rank tests ($n = 7$, $p < 0.05$), show that PolyLoss achieves the best mean AUC on Avazu (0.7675 with DNN, +0.44\% over baseline) and on Trendyol (0.6710 with DCN, +0.56\% over baseline). These results demonstrate that algorithm-level methods show greater consistency in performance improvements across diverse datasets and models, compared to data-level approaches.
Deep learning-based public transport demand forecasting
As cities grow more crowded, efficient public transport systems become more essential. The digitalization of public transport generates large amounts of data. This data creates new opportunities for data-driven optimization. One key application of such data is demand forecasting, which is essential for effective planning and resource allocation. However, traditional methods struggle to handle the complexity of urban transit data. As a result, recent research has focused on deep learning models which achieve higher accuracy by capturing complex temporal and spatial patterns. This thesis introduces Mini-PatchTST, a lightweight variant of PatchTST, designed for metro passenger demand forecasting with reduced complexity. To support this work, a novel station-level hourly metro passenger flow dataset was constructed using open data from the Istanbul Metropolitan Municipality. The proposed approach is bench- marked against traditional statistical methods, machine learning models such as sta- tistical methods, machine learning models (LightGBM, XGBoost), and deep learning baselines (LSTM, PatchTST). Experimental results on the Istanbul Metro dataset show that Mini-PatchTST outper- forms gradient-boosting models with a 5–13% RMSE reduction while maintaining com- parable MAE. It achieves a MAPE of 23.42%, substantially improving upon PatchTST (33.89%) though remaining slightly higher than tree-based models (15–16%). Compared to the original PatchTST, Mini-PatchTST reduces RMSE by 4%, MAE by 10%, and MAPE by nearly 31%. Although LSTM attains the lowest RMSE and MAE, its higher MAPE and significantly longer training time highlight the superior efficiency–accuracy trade-off of Mini-PatchTST. To assess generalization capability, the model is further evaluated on the open-source Hangzhou Metro dataset, where it yields substantial im- provements over models used in prior studies. The findings indicate that Mini-PatchTST provides a robust, and computationally efficient solution for short-term public transport demand forecasting.
AI-Based cost estimation for automotive suspension and steering parts
This thesis presents a machine learning approach that predicts unit manufacturing cost directly from two dimensional engineering drawings of suspension and steering parts. The dataset contains 13,684 drawings grouped into twenty four product categories. Drawings are parsed to extract geometric and dimensional entities, and about two hundred features are derived for each case. These features include descriptive statistics for lines, arcs, circles, and dimension values, histogram summaries with twelve bins, and distances between drawing level histograms and product group references using the Euclidean distance and the Kullback--Leibler divergence. Cost is modeled with methods that use gradient boosting with decision trees. Three machine learning algorithms are evaluated, namely XGBoost, CatBoost, and LightGBM. Models are trained per product group under a common validation and tuning protocol. Across the twenty four groups, the average mean absolute percentage error (MAPE) for XGBoost, CatBoost, and LightGBM is about 11% on the test dataset. Interpretation through feature importance and SHAP shows that maxima of rotated dimensions, statistics that describe arc geometry, and divergence measures from the histogram comparisons are consistent drivers of cost. The proposed end-to-end pipeline has the potential to operate without detailed process plans, shorten quotation lead times, improve consistency within part families, and be linked to design or enterprise systems to provide near real-time cost feedback.
An al-based approach to academic quality quantification
In the context of academic recruitment at universities and research institutions, establish- ing consistent and effective evaluation criteria remains a complex challenge. Identifying robust metrics that align with globally recognized standards of academic quality is essen- tial to ensure merit-based evaluation of researchers and maintain institutional credibility. In this thesis, we address this challenge through an AI-based solution aimed at developing a data-driven approach to quantify academic quality. As a benchmark of academic excellence, we use Nobel laureates' profiles in Physics, Chemistry, Physiology or Medicine, and Economics as a reference cohort. Comparison group includes researchers from the same fields affiliated with universities with an average ranking based on the Times Higher Education World University Rankings. By defining bibliometric features of the academic profiles of both Nobel laureates and the comparison group, we aim to develop machine learning models to quantify academic quality and identify the key features that define academic excellence. The ultimate goal is to support the decision-making process of universities and research institutions in academic recruitment, creating a fair and objective evaluation criteria.