Web search engines allow users to find information on almost any topic imaginable. To be successful, a search engine must return relevant information to the user in a short amount of time. However, efficiency (speed) and effectiveness (relevance) are competing forces that often counteract each other. It is often the case that methods developed for improving effectiveness incur moderate-to-large computational costs, thus sustained effectiveness gains typically have to be counter-balanced by buying more/faster hardware, implementing caching strategies if possible, or spending additional effort in low-level optimizations.

This thesis describes the "Learning to Efficiently Rank" framework for building highly effective ranking models for Web-scale data, without sacrificing run-time efficiency for returning results. It introduces new classes of ranking models that have the capability of being simultaneously fast and effective, and discusses the issue of how to optimize the models for speed and effectiveness. More specifically, a series of concrete instantiations of the general "Learning to Efficiently Rank" framework are illustrated in detail. First, given a desired tradeoff between effectiveness/efficiency, efficient linear models, which have a mechanism to directly optimize the tradeoff metric and achieve an optimal balance between effectiveness/efficiency, are introduced. Second, temporally constrained models for returning the most effective ranked results possible under a time constraint are described. Third, a cascade ranking model for efficient top-K retrieval over Web-scale documents is proposed, where the ranking effectiveness and efficiency are simultaneously optimized. Finally, a constrained cascade for returning results within time constraints by simultaneously reducing document set size and unnecessary features is discussed in detail.