Activities Prediction Based On Spatiotemporal Semantic Association

With the rapid development of world economy and technology,many Geo-Tagged Social Media(GTSM)have risen rapidly,resulting in a large number of spatiotemporal text semantic data with hot activity tags.These data have instructive significance for predicting people's activities.This thesis takes spatiotemporal semantic activities prediction as a research problem,and focuses on the prediction methods of temporal and spatial semantic activities.The main contents are as follows:(1)Spatiotemporal semantic activity prediction based on Skip GRUThe traditional multimodal embedding spatiotemporal semantic activities prediction method treats the text as a vocabulary,and loses the sequence of the text.In order to solve this problem and improve the prediction accuracy,this thesis proposes a method of spatiotemporal semantic activities prediction based on Skip GRU(SAPSG).Firstly,the shared parameter matrix of GRU(Gated Recurrent Unit)is used to recursively extract text sequence features along the text word order direction.Secondly,in order to improve the convergence rate of the activity prediction model,this article adds the Skip mechanism to the GRU.The Skip mechanism adds a parameter update penalty mechanism on the basis of the original GRU neurons to suppress the redundant update of the neuron state,thereby speeding up the model convergence.The experimental results show that SAPSG has increased the accuracy by 3.06% compared with Re Act method,and the convergence rate of the model has increased by 26.8%compared with GRU.(2)Spatiotemporal semantic activity prediction based on ensemble SkipTinyGRUThe prediction activity of a single model is prone to overfitting,resulting in low prediction accuracy,and the numerous parameters of Skip GRU will cause the model to occupy a large amount of memory.In order to solve the above problems,this thesis proposes a spatiotemporal semantic activity prediction method(SAPEST)based on integrated SkipTinyGRU.Firstly,optimize the activity prediction model from the starting point of multiple parameters to reduce the risk of overfitting.Secondly,add the Tiny mechanism to Skip GRU.The Tiny mechanism uses the weighted residuals and the sparse low-rank representation of the neuron parameter matrix to reduce the model footprint.The experimental results show that the accuracy of SAPEST method is 0.57%higher than that of SAPSG method,and the stock is reduced by 5.7%.(3)Activities prediction based on ensemble SkipTinyGRU and spatiotemporal semantic graph convolutionThe aforementioned method only considers the characteristics of a single activity point,and does not take into account the spatiotemporal semantic associations between different activity points,resulting in a decrease in the accuracy of activity prediction.In order to solve the above problems,this thesis proposes an activity prediction method(APESGC)based on ensemble SkipTinyGRU and spatiotemporal semantic graph convolution.Firstly,by comprehensively considering the temporal and spatial semantic similarity,time interval and distance interval between the activity points,the adjacency matrix of the temporal and spatial semantic activity graph is established.Secondly,the integrated SkipTinyGRU is used to extract the features of the active points,and the sum of the output vectors of the fully connected layer of each base learner is used as the initial vector of the corresponding active points in the activity map.Finally,use graph convolutional neural network to predict activity.Experimental results show that the accuracy rate of APESGC is 1.42% higher than that of SAPEST.The thesis has 26 figures,11 tables,and 98 references.