Jiaxin Huang, Chunyuan Li, Krishan Subudhi, Damien Jose, Shobana Balakrishnan, Weizhu Chen, Baolin Peng, Jianfeng Gao, Jiawei Han. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021.

In web search, typically a candidate generation step selects a small set of documents---from collections containing as many as billions of web pages---that are subsequently ranked and pruned before being presented to the user. In Bing, the candidate generation involves scanning the index using statically designed match plans that prescribe sequences of different match criteria and stopping conditions. In this work, we pose match planning as a reinforcement learning task and observe up to 20% reduction in index blocks accessed, with small or no degradation in the quality of the candidate sets.

This paper studies the evaluation of policies that recommend an ordered set of items (e.g., a ranking) based on some context---a common scenario in web search, ads, and recommendation. We build on techniques from combinatorial bandits to introduce a new practical estimator that uses logged data to estimate a policy's performance. A thorough empirical evaluation on real-world data reveals that our estimator is accurate in a variety of settings, including as a subroutine in a learning-to-rank task, where it achieves competitive performance. We derive conditions under which our estimator is unbiased---these conditions are weaker than prior heuristics for slate evaluation---and experimentally demonstrate a smaller bias than parametric approaches, even when these conditions are violated. Finally, our theory and experiments also show exponential savings in the amount of required data compared with general unbiased estimators.

Ranking documents using their historical click-through rate (CTR) can improve relevance for frequently occurring queries, i.e., so-called head queries. It is difficult to use such click signals on non-head queries as they receive fewer clicks. In this paper, we address the challenge of dealing with torso queries on which the production ranker is performing poorly. Torso queries are queries that occur frequently enough so that they are not considered as tail queries and yet not frequently enough to be head queries either. They comprise a large portion of most commercial search engines' traffic, so the presence of a large number of underperforming torso queries can harm the overall performance significantly. We propose a practical method for dealing with such cases, drawing inspiration from the literature on learning to rank (LTR). Our method requires relatively few clicks from users to derive a strong re-ranking signal by comparing document relevance between pairs of documents instead of using absolute numbers of clicks per document. By infusing a modest amount of exploration into the ranked lists produced by a production ranker and extracting preferences between documents, we obtain substantial improvements over the production ranker in terms of page-level online metrics. We use an exploration dataset consisting of real user clicks from a large-scale commercial search engine to demonstrate the effectiveness of the method. We conduct further experimentation on public benchmark data using simulated clicks to gain insight into the inner workings of the proposed method. Our results indicate a need for LTR methods that make more explicit use of the query and other contextual information.