Research On Graph Representation Learning And Prediction Of Multi-Dimensional Time Series

With the rapid development of information technology,massive amounts of data has been generated in global industrial productions,economic activities,transportation and other fields.Among them,time series data is an ordered collection of data collected in chronological order,which contains a wealth of information behind it.The predictions of time series values,trends and fluctuations have important application value and theoretical significance.In recent years,With the development of research of artificial intelligence,the intelligent models for the analysis and prediction of time series data have received wide attention.Different from the traditional econometric models,machine learning models show strong learning ability for nonlinear,nonstationary and highly noisy time series data,but there is still more space to improve their interpretability and generalizability.At the same time,researches using graph learning methods on the analysis of time series data such as spatio-temporal,multi-dimensional,etc.,has become increasingly active.According to these researches,how to make effective graph feature representation of multi-dimensional time series is the primary work and necessary condition for adopting graph learning methods.In this thesis,we explore the research areas of graph representation learning and forecasting of time series,the main research contents are as follows.(1)A time series prediction model with multi-reservoir fuzzy cognitive map is proposed to enhance the interpretability and generalizability of the time series prediction model.Here,the concept nodes in the fuzzy cognitive graph are replaced by echo state network modules,which in turn are used to learn the temporal state features from the sequence data.The model has the advantages of fuzzy cognitive graph causal inference and interpretability,as well as the ability of echo state networks to capture long-term temporal dependence.The experimental results verify that the designed model has high prediction accuracy and good interpretability.(2)A type of time-frequency contrastive learning model is designed to represent the temporal dependencies of multi-dimensional time series and multi-series data correlations in both time and frequency domains in a characteristic manner,so that an effective graph structure representation of multi-dimensional time series is obtained.Firstly,Mixup data enhancement is used to obtain the enhanced data.The amplitude and phase of the original and the enhanced data are extracted separately using Fourier transform.Secondly,the time-domain and the frequency-domain loss of the original and the enhanced data are calculated by encoder separately,and the two parts of the loss are summed to calculate the total loss.Finally,the graph structure representation of the original data is obtained from the trained model.A large number of manual and real datasets evaluated to verify the validity of the obtained time series feature representations and the reasonableness of the graph structure representations.Further,by using the above graph representation method,a time series prediction model based on spatio-temporal multi-head sparse attention is designed and applied to the traffic prediction problem.Multiple sets of experimental results show that timefrequency contrastive learning can effectively achieve high-quality graph structure representation,and the better prediction performance can be obtained by migrating the graph structure representation of this time series to the downstream prediction task.