Adapting Rankers Online

At the heart of many effective approaches to the core information retrieval problem— identifying relevant content—lies the following three-fold strategy: obtaining contentbased matches, inferring additional ranking criteria and constraints, and combining all of the above so as to arrive at a single ranking of retrieval units. Over the years, many models have been proposed for content-based matching, with particular attention being paid to estimations of query models and document models. Different task and user scenarios have given rise to the study and use of priors and noncontent-based ranking criteria such as freshness, authoritativeness, and credibility. The issue of search result combinations, whether ranked-based, score-based or both, has been a recurring theme for many years. As retrieval systems become more complex, learning to rank approaches are being developed to automatically tune the parameters for integrating multiple ways of ranking documents. This is the issue on which we will focus in the talk. Search engines are typically tuned offline; they are tuned manually or using machine learning methods to fit a specific search environment. These efforts require substantial human resources and are therefore only economical for relatively large groups of users and search environments. More importantly, they are inherently static and disregard the dynamic nature of search environments, where collections change and users acquire knowledge and adapt their search behaviors. Using online learning to rank approaches, retrieval systems can learn directly from implicit feedback, while they are running. The talk will discuss three issues around online learning to rank: balancing exploitation and exploration, gathering data using one pair of rankers and using it to compare another pair of rankers, and the use of rich contextual data.