Location Prediction In Social Networks

User locations in social networks are needed in many applications which utilize location information to recommend local news and places of interest to users,as well as detect and alert emergencies around users.However,considering individual privacy,only a small potion users share their locations on social networks.Therefore,in order to obtain more accurate locations of users to provide services to address-based applications,this thesis proposes an address prediction framework.The framework constructs corresponding models from three different perspectives,namely,content-based model,timing-based model and joint model which combine these two models.In the content-based model,we filter out those location-independent tweets and use deep learning algorithm to mine the relationship between semantics and locations.In the timing-based model,users' daily behaviors habits are mined by using historical data,and users' behavior rules are analyzed to predict their current position.The first two models are fused based on the timing and text features,and location information is obtained from these two features.In this thesis,our main research work divided into the following four parts.(i)We filter out micro-blogs that do not contain any location related words from raw dataset.And,the concept of location related words is defined as words with strong relevance to locations.These words are found out through keyword extraction algorithm.(ii)We conduct data processing on the content of users' microblog posts,including word segmentation,and stop word removal.Firstly,by improving the existing word vector model,we trained word vector model to transform word segmentation data.The address tag corresponding to the microblog is mapped to one of the grids through the city grid,and the grid number becomes the "new" address tag.(iii)Aiming at the above filtered and preprocessed data sets,we construct convolution neural network and cascade LSTM network based on text and time series respectively.And we mine these two features at the same time,merge ConvLSTM network.We also improved the activation function and loss function of basic LSTM.(iv)Finally,we test and analyze the effect of the proposed location prediction model,and compare it with the related models to verify the effectiveness of the model.The we tested and analyzed the LSTM network with improved CBOW model,replacement activation function and improved loss function.The prediction effect of the improved models are better than the basic network.