Hila Becker, Christopher Meek, David Maxwell Chickering
In this paper, we develop and evaluate several probabilistic models of user click-through behavior that are appropriate for modeling the click-through rates of items that are presented to the user in a list. Potential applications include modeling the click-through rates of search results from a search engine, items ranked by a recommendation system, and search advertisements returned by a search engine. Our models capture contextual factors related to the presentation as well as the underlying relevance or quality of the item. We focus on two types of contextual factors for a given item; the positional context of the item and the quality of the other results. We evaluate our models on a search advertising dataset from Microsoft's Live search engine and demonstrate that modeling contextual factors improves the accuracy of click-through models.
Subjects: 12. Machine Learning and Discovery; 1. Applications
Submitted: Apr 24, 2007