| Online social networks have changed people’s way of life and become the main way for user communication,and the resulting large amounts of user behavior data can be used to identify user preferences and improve personalized recommendation performance.Location-based social networks further expand users’ online interaction behaviors to the offline world and gather many users spatio-temporal information.These multi-level and multi-angle heterogeneous data reflect the characteristics of user interaction and user mobility,which brings opportunities to promote the construction of more accurate and effective recommendation systems.On the other hand,it also makes recommendation methods face more significant challenges in how identifying useful features from the massive and sparse data and how to realize the integration of multi-dimension data.Based on users’ social and trajectory data,the study of personalized recommendation integrating multiple heterogeneous data is not only of theoretical value in understanding users’ behavior and improving the effectiveness of personalized recommendation algorithms but also of practical significance in guiding enterprises to improve their service level and expand their business.This thesis takes location-based social networks as the background and investigates the problem of personalized recommendation integrating users’ multidimensional data by analyzing users’ online social relationships,offline location interaction,and user mobility regularity.For each recommendation task,to improve the recommendation effect,effective features are extracted from multidimensional data,and graph neural network models are constructed to integrate the information based on distinguishing the influence of different features.Specifically,the main research problems and innovations of this thesis are as follows:(1)Friend recommendation in LBSNs integrating trajectory data.To identify and extract the similarity features exhibited in user trajectories and achieve effective integration of trajectory information and social information,a dual subgraph-based pairwise graph neural network is proposed in chapter 3 for friendship prediction in location-based social networks.The model first constructs a high-order location graph based on the location co-occurrence in all users’ trajectories by using an entropy-based random walk which can consider location popularity,after that,the trajectories of different user pairs can be represented as subgraphs in the location graph,based on which the subgraph-based graph neural network is used to learn user mobility similarity.Meanwhile,the social proximity of user pairs can be learned by a graph neural network model based on social network subgraphs.Finally,the adaptive fusion of user social proximity and mobility similarity is achieved through the gating layer for relationship prediction between users.The experimental results on three real datasets show that DSGNN can effectively identify the user similarity in their trajectories,while achieving an effective fusion of social information and trajectory information.(2)Location recommendation in LBSNs integrating trajectory data.The trajectory represents the spatiotemporal regularity of users’ mobility,the location semantics represents the topic of users’ activities,and the similarity of location preferences among friends can mitigate the negative impact of sparse user check-ins.To fuse different contextual information,a multi-context-based next location recommendation model is presented in chapter 4.By analyzing the spatio-temporal correlation,the model can identify the most relevant historical trajectory sequences for users’ current conditions for user preference learning.Meanwhile,the model improves target user preference prediction by identifying the friends with consistent preferences from user trajectories.The sequence relationship and semantic relationship between locations are also analyzed to capture multiple correlations of the location to improve users’ short-term preference modeling.Finally,the modeling of user preferences is achieved by aggregating multiple contextual information.The experimental results on three real datasets show that the model can effectively improve the next location recommendation effect compared with other methods.(3)Travel companion recommendation in LBSNs integrating trajectory data.With the development of social networks,different user interaction content and interaction methods have been generated,online companion searching behavior becomes more and more common,i.e.,by posting travel plans on the platform to look for users who can travel together.Facing this new recommendation scenario,a two-stage companion recommendation model is proposed in chapter 5 for travel companion recommendation,which can recommend users who not only have the same interest in the travel destinations but also have a high tendency to be friends with the target users.The model first identifies different users’ preferences for the destination according to their trajectory information and obtains user preference representations,then analyzes different users’ companionship tendencies from the perspective of their personal traits using two indicators: companionship rate and desire to explore,finally,uses a subgraph-based neural network framework to integrate user preferences,personal traits and social intimacy features to filter and recommend travel companions.The experimental results show that the model can effectively analyze the social closeness,companionship tendency,and preference consistency among users and achieve efficient travel companion recommendations.This thesis analyzes the role of social information and trajectory information in different recommendation problems of location-based social networks and proposes graph neural network-based models to identify,extract,and integrate different features to improve the effectiveness of different personalized recommendation tasks.The research results enrich the research about user behavior analysis and feature mining in social networks,provide new ideas for personalized recommendation methods that integrate heterogeneous data,and provide references for social platforms and enterprises to identify user preferences based on user behavior mining,expand platform business,and promote enterprise growth. |
