Research On Tourism Personalized Location-based Service Integrating Multi-Source Information

Tourism industry plays an important role in promoting economy growth and improving people's quality of life.With the development of China's economy and the improvement of people's living conditions,China's tourism industry has entered a period of rapid development.As the popularization of the mobile internet,and the development of the location-aware technologies and the big data technologies,increasingly people hope to access online personalized tourism information and services during the tourism process,e.g.,online booking tickets of scenic spots,online searching the nearby hotels and restaurants and online planning itineraries.The key to realize the above smart services is to design and implement the tourism oriented techniques and methods of personalized location service.However,due to the relatively under research on personalized location services for tourism or urban travel,the problems of low quality of tourism services and tourism information overload have been caused to tourists,which limit tourism experience and affect the development of tourism-related industries.Due to the vast economic and social value of the tourism personalized location service technology,this research domain has attracted much attention of the industry and academia recently.In real tourism activities,the personalized location services are mainly divided into two types: travel route recommendation and point of interest(POI)recommendation.However,there are still several research challenges and problems in implementing these two services.First,the real tourism behavior of tourists is rich but difficult to obtain,which limits the accuracy of the systems in learning tourists' preferences.Second,the existing travel route recommendation methods are hard to comprehensively consider the tourists' complex and various individual trip context constraints,resulting in the low personalization of the recommended routes.In addition,in POI recommendation tasks,multi-source heterogeneous tourism data is difficult to be fused and represented,resulting in the recommendation algorithms is unable to learn the comprehensive characteristics of tourists' preferences and POI tourism attributes,which limits the accuracy of the recommendation results.Finally,as the tourists' visiting behaviors in the POI recommendation tasks are relatively sparser,it is difficult for traditional personalized recommendation methods to effectively learn the high-order interactive features between tourists and POIs in the case of sparse tourist data or cold startup,which limits the relevance and personalization of the recommendation results.To overcome the above challenges,this thesis researches the following aspects in-depth: the tourist behavior awareness and acquisition technology,the travel route mining and personalized recommendation algorithms based on multiple trip constraints,the multiple tourism contexts uniformly modeling method in POI recommendation,and the deep POI deep recommendation model based on multi-knowledge joint representation.The main work and contributions of this thesis are as follows:1.The on-site tour behavior of tourists contains their personal preference and characteristics.However,due to the lack of on-site tour behavior data of tourists in specific scenic spots,the existing travel route recommendation systems are difficult to recommend tangible travel routes for tourists in specific scenic spots which comply with their personal preferences and travel constraints.To this end,a novel tangible travel route recommendation system is proposed whereby studying the on-site tourist travel behavior data acquisition and mining technologies.In detail,we first design an on-site tourist travel behavior acquisition method based on Internet of Things technology and smart phones,and to acquire fine-grained tourists' preferences by mining these on-site behavior data.Second,we propose the Tourist-Behavior pattern sequence mining algorithm,i.e.,TB-Prefix Span to generate massive candidate travel routes.Last,we design a personalized route ranking and recommending method,which takes the route value,the touring duration and the reasonability of visiting sequence into consideration concurrently.The experimental results in real-world travel circumstance demonstrate that our system and related methods are capable of recommending tangible travel routes which match tourists' personal preferences and travel constraints with respect to a specific scenic spot.2.The existing POI route recommendation systems are hard to generate high personalized and reasonable travel routes due to neglecting the rich tourism contexts and constraints during the route mining process.To this end,we propose a POI route recommendation system based on heterogeneous tourism big data.Specifically,we first design the POI knowledgebase and massive structured POI visit sequence constructing methods based on multiple heterogeneous tourism data,such as real electronic map data,travelogues of OTAs and POI attributes data.Then,the POI-Visit pattern sequence mining algorithm PV-Prefix Span is proposed to generate frequent POI travel routes.Last,by incorporating multiple tourism contexts,including the trip duration,the companion type,the visit season and the preferring tourism type,we design a personalized POI route ranking and recommending methods.The effectiveness of our system and related methods is verified by the experimental results on real tourism data set.3.In POI recommendation tasks,the personalization of recommendation results is determined by the quality of tourists' and POIs' latent feature representations learned by recommendation algorithms.However,due to the single spatiotemporal semantics contained in tourist's trajectories,it is difficult for existing works to consider other heterogeneous tourism contexts concurrently,e.g.,personal trip constraints and POI tourism attribute.To this end,we propose the Neural Multi-context Modeling Framework,NMMF,to learning tourist and POI feature representations from multiple tourism contexts.Specifically,regarding the graphic structure POI attribute context and the sequential structure tourist trajectory context,we design TKG2 vec and Traj2 vec model respectively to learn POI tourism attribute and tourist visit sequential feature vectors.Then,a feature representation integration strategy is designed to integrate feature vectors learned by two models,so as to construct comprehensive tourist and POI feature representations.The experimental results on real tourism data set show that NMMF outperforms two kinds of the context modeling based POI recommendation methods.4.How to model user and item feature representations from interactive data is the key point of personalized recommendation research.However,the tourist historical visit behavior data is relatively sparse in the POI recommendation,which makes traditional recommendation algorithms failing to be directly applied to this task.To this end,on the basis of studying the methods of incorporating multi-knowledge into the personalized POI recommendations,we design the Multi-Knowledge Embedding POI Recommendation Model,abbr.MKERM.In detail,we first design two knowledge pre-train models to uniformly represent POI attribute knowledge and user behavior sequence knowledge.Then,we design the Multiple Knowledge Graph Convolutional Network,MKGCN,and the User Preference Dynamic Encoding Network,UPDEN,to construct POI and user comprehensive semantic feature representations,respectively.Finally,we design the objective function of MKERM to learn the partial ordering relation of positive and negative POI pairs,which enhances the capabilitiy of MKERM in capturing high order interative features of user and POI.The experimental results prove that MKERM outperforms the baseline methods in the case of sparse user data.