Deep Meta Learning For Predictions Of Urban Spatio-Temporal Data

In the big data era,data has become an important part in the develop-ment of technology and society.In recent years,with the advance of mobile sensing techniques,large amounts of spatio-temporal data have been col-lected,and more and more companies and governments realized the values of these data.Different from the conventional data types,these data consist of the static spatial properties and dynamic temporal readings of the loca-tions all over a city,capable of reflecting the trends of the city states.Thus,predicting these spatio-temporal data is essential to help us solve many prob-lems in the modern city,including traffic congestion,air pollution,energy consumption,and etc,which can improve the efficiency and convenience of a city.Therefore,accurately predicting spatio-temporal data is an important step to build a smart city.In the field of spatio-temporal data prediction,many works have been proposed.In the beginning,most of works adopt traditional machine learn-ing techniques.However,as these methods have low model complexity,they require large amounts of human crafted features to make accurate predic-tions.Thus,these traditional machine learning methods cannot generalize to a wide range of spatio-temporal prediction tasks.Recently,with the ad-vances of large-scale computing infrastructures and collections of massive data,deep learning techniques have become a new perspective in spatio-temporal data mining.Currently,deep learning based methods employ many network structures to learn spatio-temporal correlations from data,such as convolution neural network,recurrent neural network,and etc.However,the existing deep learning models ignore the generation pro-cess of spatio-temporal correlations.Thus,in this thesis,we aim to study this process and its impact.To this end,we propose several problems,including the diversity of spatio-temporal correlations,the diversity of spatio-temporal correlations in spatial domain,the diversity of spatio-temporal correlations in temporal domain,and the automatic spatio-temporal neural architecture design.To tackle the above problems,we propose several frameworks re-spectively,such that the predictions of spatio-temporal data can be improved.First,to capture diverse spatio-temporal correlations,we propose a deep learning framework,entitled MF-STN,to enhance the state-of-the-art deep spatio-temporal models.In particular,we use a matrix factorization technique on neural networks,namely,decomposing the region-specific pa-rameters of the predictor into learnable matrices,i.e.,region embedding matrices and parameter embedding matrices,such that the latent region function along with the diverse spatio-temporal correlations among regions can be modeled.Second,to capture the relationships between spatial information and spatio-temporal correlations,we proposed a deep meta learning framework,entitled ST-Meta Net,to collectively predict traffic in all locations at once.This framework learns the meta knowledge from spatial information,and then uses such meta knowledge to generate the parameter weights in spatio-temporal neural networks.In this way,the relationships between spatial information and spatio-temporal correlations can be captured.Third,to further capture the relationships between temporal information and spatio-temporal correlations,we improve ST-Meta Net,and propose a new framework ST-Meta Net~+.This framework learns the temporal context from temporal information,fuses it with the meta knowledge learned from spatial information,and then generates the parameter weights in spatio-temporal neural networks.In this way,the relationships between temporal information and spatio-temporal correlations can be modeled in advance.At last,to automatically design spatio-temporal neural networks,we propose a neural architecture search framework,entitled ST-NAS.We design the search space with a candidate operation set for modeling spatio-temporal correlations,and then employ a neural architecture search algorithm to find possible neural networks.In this framework,we also employ a weight generation based method to tackle diversity of spatio-temporal correlations.In this thesis,we evaluate all the above frameworks by several flow prediction and traffic speed prediction tasks,to show the advantage of our proposed frameworks and illustrate our contributions.