Representation Learning And Prediction Modeling For Spatial-temporal Data

With the continuous upgrading of national intelligent construction strategy and the fast development of techniques such as real-time positioning,intelligent mobile terminals and portable sensors,spatial-temporal data has become increasingly available nowadays and shows practical application value.Mining valuable knowledge from spatial-temporal data can contribute to effectively solve the problems in real world applications.Accordingly,how to better learn the representation of pattern in spatial-temporal data,and establish accurate prediction model for different tasks,so as to realize discovering of highvalue knowledge automatically can be a very valuable research problem.In this thesis,spatial-temporal data is taken as the research object,focusing on the key research task of representation learning and prediction modeling in spatial-temporal data mining,aiming at the important research issues such as taking importance-based sampling for time series prediction,heterogeneous spatial-temporal representation learning for demand prediction,learning embedding from order-dependently semantic cross-field categorical attributes and dynamic interest sequence learning via knowledge graph.The contributions of this thesis are summarized as follows.(1)An evolutionary attention-based LSTM,named EA-LSTM,is proposed for enhancing the ability to importance-based sampling during temporal relationship mining in traditional LSTM-based models.To refrain from being trapped into partial optimization like traditional gradient-based methods,an evolutionary computation inspired competitive random search method is also proposed,which can well configure the parameters in the attention layer of EA-LSTM model.Experimental results on different time series prediction task have illustrated that the proposed model can achieve significantly better prediction performance compared with other baselines.(2)In bike-sharing demand prediction task,a dynamic heterogeneous graph is constructed by complex spatial-temporal relations among user,riding tips and bike stations and a novel model named STG2 Vec is proposed to learn the representation from heterogeneous spatial-temporal graph.Additionally,together with other multi-source information such as geographical position,historical transition patterns and weather,e.g.,the representation learned by STG2 Vec can be fed into the LSTMs for make a station-level bike-sharing demand prediction.Experimental results on Citi-bike dataset have illustrated that the proposed model can achieve significantly better demand prediction performance compared with other baselines.(3)A cross-field categorical attribute embedding model(CCAE)is proposed for the widely existing categorical structured data in EMR.Through the order-by-order feature interaction,the vectorized representation in attribute-level by orders,in which the strong semantic coupling among categorical variables can be well exploited.Furthermore,by transforming the order-dependency modeling into a sequence learning task in an ingenious way,Recurrent Neural Network is adopted to capture the semantic relevance among multi-order representations.Experimental results on SEER EHR dataset have illustrated that the CCAE can achieve better clinic endpoint prediction performance compared with other baselines.(4)In click through-rate prediction task,to address the issue of lack of modeling user's dynamic interest in exiting path-based and embedding-based models with knowledge graphs,a dynamic interest sequence learning(DISL)method is proposed.Specifically,a multi-granularity dynamic interest sequence learning method,which is based on knowledge-enhanced path mining and interest fluctuation signal discovery is proposed to model user's interest preference from his historical click behaviors.Experimental results on different datasets have illustrated that the proposed method can achieve better recommendation performance compared with other base-lines...