Advancing neural click models for unbiased learning-to-rank

Search engines and recommender systems often rely on implicit user feedback, particularly clicks, to optimize their ranking algorithms. Clicks, however, are a biased signal of user preference. Users can click only the items shown to them and tend to inspect top-ranked items more thoroughly, leading to fewer clicks on lower-ranked but relevant items. Mitigating these statistical biases in click data while optimizing ranking algorithms is the goal of unbiased learning-to-rank (ULTR). Click models are a prevalent ULTR method for bias mitigation by explicitly modeling user behavior, such as which items a user examined, clicked, or was satisfied with. In this thesis, we advance our understanding of neural click models, a modern iteration of traditional click models that integrates neural networks for better generalization, incorporates more complex user feedback, and scales to massive datasets through gradient-based optimization. Over four chapters, we investigate the theoretical properties, real-world effectiveness, and scalability of neural click models for bias correction in ranking.