Research On Personalized POI Retrieval At Scale

After years of development,Baidu Maps has realized the evolution from a trip tool to the digital base of the new infrastructure in the era of intelligent transportation by providing users with accurate,rich and intelligent personalized travel services.The daily requests of location service at Baidu Maps have exceeded 120 billion times.The number of monthly active users at Baidu Maps has exceeded 450 million.Moreover,the total number of points of interest(POIs)has reached 180 million.Personalized POI retrieval is designed to allow hundreds of millions of map users to quickly and accurately find the required POIs at anytime and anywhere from the largescale POI database.It is not only an important entry for many personalized services in web mapping services,but also one of the most frequently used functions by users.Personalized POI retrieval not only needs to consider the query words or incomplete query fragments entered by the user,but also the user preference modeling,and even needs to include time and spatial information such as the time when the retrieval was initiated and the location where the retrieval occurred.For this reason,it has attracted increasing interest from both academia and industry.To this end,this thesis has carried out in-depth research on personalized POI retrieval methods and large-scale industrial applications.The research is conducted with National Key Research and Development Program(2018YFB 1004304)and the practical technical challenges of Baidu Maps.The main work and contributions of this thesis are listed as follows:1.A personalized POI retrieval method based on user’s historical behavior.This method first uses the historical data of users’ search behavior to automatically learn the relationship between the user’s personalized feature attributes and POI through semantic analysis.It establishes a personalized POI embedding approach that can characterize user preferences.Then,the learned users and POI embedding are incorporated into the POI ranking model,so that the candidate POIs can be ranked and optimized according to the preferences of each user.The experimental results on the large-scale dataset of Baidu Maps show that this method can significantly improve the effectiveness of personalized POI retrieval.2.A personalized POI retrieval method based on personalized prefix representation learning.This method maps both prefixes and candidate POIs into their corresponding vector representations in the same semantic space,and uses the semantic similarity between the two vector features as personalized features to adjust the POI ranking results.The experimental results on the large-scale real-world dataset of Baidu Maps show that this method can significantly improve the efficiency and accuracy of personalized POI retrieval.3.A personalized POI retrieval method based on spatial-temporal learning and meta-learning.This method maps time,space,user historical behaviors,input prefixes,and candidate POIs into the same semantic space for unified encoding.It uses a large-scale historical search log of users for learning,so that all personalization factors are taken into account for ranking.Experimental results on the large-scale dataset show that this method can significantly improve the overall performance of POI retrieval at Baidu Maps.In addition,we further propose a learning to rank method based on spatialtemporal modeling and meta-learning.This method can generate specific POI retrieval models for different spatial-temporal data,and can provide better POI retrieval results for queries that are issued in long-tail spatial or temporal distribution.Experimental results show that this method can further improve the user experience of personalized POI retrieval.In terms of large-scale industrial applications,the above research has been successfully deployed at Baidu Maps.The personalized POI search service responds to hundreds of millions of requests every day,saving users’ 38.5%of input time and significantly improving the search experience of Baidu Maps users.