Research On Geographic Location Positioning Based On Social Network User Content And Association Relationship

With the rapid development of online social networking platforms,a large amount of unstructured data has been generated,such as text content posted by users,topic tags participating in discussions,and mutual attention and interaction between users.Although the popularity of positioning devices makes location information easy to obtain,such sensitive data is limited to specific social platforms.Therefore,how to infer the geographic location of a user's address based on user-generated content and behavior characteristics has become the focus of attention.The determination of the user's geographic location has become the key to many downstream applications providing services,such as locationbased targeted advertising,recommendation of local events/locations,restricted content distribution following regional policies,and so on.Some existing methods can solve the problem of user location positioning.Most of them use a segmentation algorithm to divide users into different areas,and the area number is the user label.In this way,the user location positioning problem can be converted into a classification problem.Although the previous work has achieved gratifying performance,there are still four problems as follows: First,the previous method embeds the user text content as a fixed representation and cannot capture the style of the user text content.Second,the noisy information of text and networks cannot be effectively processed and filtered.Third,in terms of social network feature learning,the existing models do not make full use of the user's topological structure,lacking the mining of the features of isolated and unlabeled nodes.Furthermore,they cannot identify and apply the attributes of the population in the cluster.Fourth,the model lacks interpretability,especially for the neural network-based model.Thus they unable to analyze the key factors that affect the positioning performance of the models.This thesis proposes a Multi-aspects Attention Graph Neural Network(MAGNN)and a Hierarchical Graph Neural Network(HGNN)to solve the above problems.MAGNN unifies text content and interactive network to perform end-to-end user location prediction and dynamically learns the embedded representation of text content and network features according to specific tasks.The application of the attention mechanism makes the model have the ability to capture various information from multiple data sources,effectively capture the style of text content,which make the model has the ability to distinguish the importance of content and nodes so as to solve the interference information in the content and network structure.HGNN is a location-aware positioning method that can integrate multiple aspects of data.It combines the user's geographic location information and the attributes of the crowd gathered in the area.It can capture the topological relationship while retaining the relative position of the node and the area.HGNN uses hierarchical graph learning to encode the structure and features of the region,which effectively alleviates the fusion problem of noisy and unstable signals and makes full use of and mining user topological structure features.Besides,an association mechanism is designed in HGNN to bridge the connections between users and regions,achieve the purpose of aggregation of isolated node information,and capture the relationship between unlabeled nodes and labeled subgraphs.Finally,an influence function is introduced to explain the model's behavior by identifying the importance of data samples when predicting the user's location.This method provides meaningful explanations about model behavior and prediction results and overcomes the shortcomings of the lack of interpretability of positioning models previously regarded as “black-box”.This thesis comprehensively evaluates the model on three real Twitter datasets.The experimental results verify that the MAGNN and HGNN models have superior performance compared with the existing benchmark methods and perform interpretability analysis,which will help readers and model designers understand user location prediction's problem better.