Personalized Query Auto-Completion for Large-Scale POI Search at Baidu Maps

Abstract

Query auto-completion (QAC) is a featured function that has been widely adopted by many sub-domains of search. It can dramatically reduce the number of typed characters and avoid spelling mistakes. These merits of QAC are highlighted to improve user satisfaction, especially when users intend to type in a query on mobile devices. In this article, we will present our industrial solution to the personalized QAC for the point of interest (POI) search at Baidu Maps, a well-known Web mapping service on mobiles in China. The industrial solution makes a good tradeoff between the offline effectiveness of a novel neural learning model that we devised for feature generation and the online efficiency of an off-the-shelf learning to rank (LTR) approach for the real-time suggestion. Besides some practical lessons from how a real-world QAC system is built and deployed in Baidu Maps to facilitate a large number of users in searching tens of millions of POIs, we mainly explore two specific features for the personalized QAC function of the POI search engine: the spatial-temporal characteristics of POIs and the historically queried POIs of individual users . We leverage the large-volume POI search logs in Baidu Maps to conduct offline evaluations of our personalized QAC model measured by multiple metrics, including Mean Reciprocal Rank (MRR), Success Rate (SR), and normalized Discounted Cumulative Gain (nDCG). Extensive experimental results demonstrate that the personalized model enhanced by the proposed features can achieve substantial improvements (i.e., +3.29% MRR, +3.78% SR@1, +5.17% SR@3, +1.96% SR@5, and +3.62% nDCG@5). After deploying this upgraded model into the POI search engine at Baidu Maps for A/B testing online, we observe that some other critical indicators, such as the average number of keystrokes and the average typing speed at keystrokes in a QAC session, which are also related to user satisfaction, decrease as well by 1.37% and 1.69%, respectively. So the conclusion is that the two kinds of features contributed by us are quite helpful in personalized mapping services for industrial practice.