Learning to rank for e-commerce search

This thesis investigates the challenges of applying learning-to-rank models in product search.
It has been shown that the vocabulary gap is larger in product search, because of the limited, unstructured nature of queries and product descriptions. Therefore, we start by focusing on query-product matching based on textual data. We conduct an evaluation of state-of-the-art supervised learning-to-match models, comparing their performance in product search. Our findings identify models that balance accuracy and efficiency, providing practical insights for practitioners.
Next, we address fairness in ranking on two-sided platforms, where the goal is to satisfy both groups of product search users: consumers and providers. Since accurate exposure estimation is crucial to achieve this balance, we introduce the phenomenon of outlierness in ranking as a factor influencing the exposure-based fair ranking algorithms. Outliers are products that deviate from others in a list, due to distinct presentational features. We show empirically that these items attract more attention and can impact exposure distribution in rankings. To account for this effect, we propose a method that reduces outlierness without compromising user utility or fairness. Moreover, we introduce outlier bias as a new type of click bias and propose an outlier-aware click model to account for both outlier and position bias.
Finally, we explore how different presentational features influence attention and perception of outliers in search results. We identify user scanning patterns and determine the role of bottom-up and top-down factors in guiding attention and shaping the perception of outliers.