Research On Spatio-temporal Sequence Prediction Method Based On Graph Representation Learning

This thesis studies the complex dependency of spatio-temporal sequence data from temporal to spatial dimensions,analyzes the challenges when modeling different types of spatio-temporal sequence data,and subsequently proposes graph representation learningbased prediction models.We conduct extensive experiments on real-world datasets,and the results validate that the proposed method is capable of predicting the future state of objects or hot event efficiently.The contributions of this thesis are as follows:(1)This thesis proposes a spatio-temporal sequence prediction framework based on graph representation learning,which can capture the dynamically-changed patterns of spatio-temporal sequences.Given the correlations between observation spots and their adjacency,and the irregular distribution of points in the target space,we denote the spatiotemporal data as a graph.We utilize graph representation learning to model the dynamic dependencies between observation spots,and then predict the value of the spatio-temporal object.Based on the unified prediction framework,we can combine different characteristics of spatio-temporal sequence to build prediction model.(2)For the sparsely-distributed discrete spatio-temporal sequence,we propose a novel prediction method integrating the global spatial similarity and local spatial correlation.The long-term and short-term temporal dependencies are then modeled by utilizing periodic characteristics and multi-source time-varying factors.In the model training stage,we design a weighted loss function in which a larger weight is assigned to the imbalanced distributed category to alleviate the zero-expansion issue.Based on the results of experiments on real traffic accident datasets,the proposed method is proved to be superior to other baselines when solving the discrete spatio-temporal series prediction problem,and it can effectively address the imbalanced distribution and zero-expansion.(3)For the densely-distributed continuous spatio-temporal sequence,we propose a novel multi-step prediction method that predicts the values of multiple time steps in the future at once,and subsequently infers the long-term development trend of objects.We design a multi-step prediction model based on graph attention network,which can adaptively capture the dynamic spatial correlation.At the same time,the inflated causal convolution with the gated mechanism is used to model the long-term temporal dependency.The residual block and skip connection are adopted in the model to avoid the gradient vanishing and speed up the training process of the neural network.The empirical experimental results show that the proposed method presents great performances in dealing with the multi-step spatio-temporal sequence prediction,and effectively alleviate the cumulative error issue in prediction.