Click-through rate (CTR) prediction is a key task in online advertising, recommendation systems, and e-commerce. In his master’s thesis, Tugay Balatlı, a graduate student at Boğaziçi University’s Data Science and Artificial Intelligence (DSAI) Institute, studied imbalance-handling techniques for deep CTR prediction. The thesis was supervised by Ercan Atam and Şaziye Betül Özateş.

The thesis, titled “Learning from Rare Clicks: A Comparative Study of Imbalance-Handling Techniques for Deep CTR Prediction,” compares data-level and algorithm-level methods, including SMOTE, NearMiss, class weighting, PolyLoss, and Dice loss, with three deep learning models: DCN, DeepFM, and DNN. Experiments were conducted on two large-scale datasets: Avazu and Trendyol JFY Ads.

The results, validated through multi-seed statistical tests, show that PolyLoss provides the most consistent performance improvements across datasets and models. The findings offer practical guidance for building more accurate and scalable CTR prediction systems in online advertising and e-commerce.