Research On Human Mobility Prediction Algorithm Based On Big Data

Human mobility prediction has broad application prospects in traffic management,advertisement placement,resource allocation and other fields.Due to the rise of social platforms in recent years,user data on social media has exploded.Predicting user movements based on social media data is gaining increasing attention.This thesis mainly studies predicting the user’s next location and jointly predicting the user’s next location and time.First,in order to predict the user’s next location,this thesis combines the DensityBased Spatial Clustering Algorithm with Noise(DBSCAN)algorithm and the DeepMove model to propose a new user mobility prediction model based on social media data.The corresponding region processing module is designed.The model uses DBSCAN to cluster regional features,and adds the region as a kind of semantic information to spatiotemporal points.The model captures the complex sequential transition patterns with time-dependent and high-order properties in human mobility through multimodal embedding and recurrent neural networks.In particular,the model associates regional shifts with location shifts,enabling it to mine region shifts hidden in location shifts.Experiments are performed on multiple real datasets,and the proposed model achieves better result than other advanced models.Compared with the original DeepMove model,the accuracy is increased by about 15%,which proves the effectiveness of the method.Then,in order to jointly predict the user’s next location and time,this thesis improves the DeepJMT model and proposes a joint location and time prediction model DeepJMT-CSL.First,for the scenario where the user lacks friends information,DeepJMT-CSL designs a method to calculate the similarity between users by using the frequency distribution of different activities of the user and the frequency distribution of the user’s visit in time,and constructs the similar user set of the user.Then,the user context information and the time context information are jointly calculated using the representations of similar users and representations of all time slots.Second,a spatial semantic module is designed to extract the user’s personalized spatial semantic information.Third,a convolution sequence module based on separable and shared convolution and dilated convolution is designed to enhance the joint extraction of nontemporal information.DeepJMT-CSL is experimentally validated on a real social media dataset.Compared with other advanced models,the proposed model achieves the best results,and its location prediction accuracy is improved by about 28%compared with DeepJMT.In addition,a joint prediction model based on location and time is also used to predict the crowd flow at the micro level.Compared with advanced macro-level crowd flow prediction models,the method shows better results.