Research On Geolocation Of Social Network Users Based On Generated Content

Over the past decade or so,social media platforms have grown rapidly,attracting large numbers of users and generating large amounts of unstructured data.One of the most useful types of information is the user’s geolocation information.Being able to pinpoint a user’s location is an important factor in many online services,such as location-based targeted advertising and local food recommendations.While internet service providers can obtain a user’s location directly from an IP address or GPS signal,such private information is not directly available to third parties.So researchers are trying to mine all kinds of information from user-generated data to locating users.Methods already in existence can be employed to address the user geolocation issue.Most of them divide the user into distinct areas through a segmentation algorithm,transforming the location issue into a categorization problem.However,there still exists some shortcomings.Firstly,the fixed embedding representation of user’s text content cannot capture the style of content.Secondly,the importance of different hop neighbor nodes to oneself is ignored in the process of feature information mixing and network representation learning.Thirdly,the influence relationship between users’ features is not considered when those various characteristic information features are mixed.Fourthly,most related works use word-level embedding for presentation learning of text content,which is inadequate for effectively filtering noise information.In conclusion,the main idea of this thesis is summarized as follows:(1)For the first two of the above shortcomings,we proposed an attention-aggregated graph convolution networks prediction model.It inherits the advantages of previous work,that is,it uses both user text content and social network relationship for modeling.Among them,attention mechanism is introduced into text content learning to effectively capture the text content style.Then,in the mixed learning of social networks and text content,an attention-aggregated graph convolution network layer is constructed for feature extraction,so that the model can distinguish the importance of multi-hop neighbor nodes of a single user node,so as to obtain more features.(2)For the last two of the above shortcomings,we proposed a metadata-enhanced graph attention networks prediction model.This model not only utilizes user text content and social network relationship,but also introduces some metadata information,that is,personal description field in user profile and post time of social dynamic.The model uses convolution neural network to learning embedded representation of text content from character-level,which can effectively filter out the noise in the text,and obtain the local features of the text content.Then,combining with the bidirectional gated recurrent unit and the attention mechanism,it can achieve the capture of globally key information in the text content,and obtain deeper feature representation.After that,it adds a graph attention neural network for training,which can attach importance to the interrelation between users’ multi-aspect features and carry out effective feature mixing.Finally,the performance of the two models proposed in this thesis is evaluated on three public datasets of real online social platforms.The experiment results show that the prediction accuracy of the proposed model is up to 9% higher than that of the latest geolocation methods,and then the performance of two models are further analyzed.