Urban Cross-domain Data Fusion Based On Computational Intelligence Technologies

This thesis focuses on investigating key problems when applying computational intelligence technologies,including neural networks and evolutionary algorithm,in the field of urban crossdomain data fusion.First,we propose to summarize the urban cross-domain data fusion scenarios by four categories,i.e.,parallel relationship,mutual relationship,causal relationship,and dynamic relationship.Specifically,the parallel data fusion category describes the scenario that multiple datasets belong to one object and these datasets are with no correlation or only week correlation.The mutual data fusion category represents the scenario that multiple datasets of one object have mutual correlations.The causal data fusion category implies the scenario that an object's one or some datasets can determine its other datasets.In addition,in some scenarios,data changes over time,which is the dynamic data fusion category.Thereafter,we study key issues of each category based on real-world applications in the field of urban economy and transportation.The main contributions of the thesis are as follows:? Aiming at investigating the fusion method for the parallel relationship-based data,we propose a two-stage-based solution that consists of two deep learning models and a light-weight prediction models.To exploit the application of the proposed solution,we apply based on a box office prediction task.Predicting box office revenue of movies before releasing on big screens successfully is of great economic importance,but very challenging as it is affected by a lot of complex factors.Owing to the limitation of the number of labeled movies,the task is a typical parallel data fusion scenario.To manage the problem,the solution extracts features of each impact factors by using these deep models in the first stage.Thereafter,taking advantage of the learned features,the box office is unlocked by the simply prediction models.In particular,the two devised deep learning model are as follows: 1)a novel dynamic heterogeneous network embedding model to simultaneously learn latent representations of actors,directors,and companies,capable of capturing their cooperation relationship collectively;2)a deep neural network-based model designed to uncover high-level representations of movie quality from trailers.Finally,we apply the solution to the Chinese film market and conduct a comprehensive performance evaluation using real-world data.Experimental results demonstrate the superior performance of both extracted knowledge and the prediction results.We have also deployed the solution as a service for a business group to assist their investment.? Aim at investigating the fusion method for mutual relationship-based data,we propose a novel evolutionary algorithm-based framework.To further exploit the application of the proposed algorithm,we apply it based on a food package suggestion problem.Ordering dishes in a restaurant is a significant task,which determines not only the customers' dining experience,but also the restaurant's reputation.However,assisting customers in ordering a satisfying food package(FP),i.e.,a combination of dishes,remains a challenge.First,local restaurants usually have very limited information about their customers,except the number of customers and their budget.Thus,suggesting FPs that satisfy their budget as well as surprise their palate is very difficult.Second,as a real-world function,FPs are required to be generated in real time while addressing several realistic issues such as dynamic dish inventories.In this study,we first extract knowledge from the history of orders of a restaurant,such as distributions of dish numbers and correlations among dishes,to formulate the FP suggestion as a multi-objective optimization problem that is a typical mutual relationship fusion scenario.Thereafter,we propose a knowledge-based multi-objective evolutionary algorithm(k-MOEA)to tackle the problem.In addition,we develop an intelligent dish-ordering system(i Ordering),including several designed online and offline mechanisms to meet the real-time requirements of the FP suggestion services.Finally,the effectiveness of the k-MOEA is evaluated quantitatively by comparing it with three categories of baselines.Moreover,we have deployed the i Ordering system in a hot pot restaurant chain,and a real-world experiment demonstrates the advanced user experience of the devised system.? Aiming at investigating the fusion method for causal relationship-based data,we propose a novel deep learning-based model.To further exploit the application of the proposed model,we apply it based on a city's rental house suggestion application.With the rapid progress of urbanization,more and more migrants need to rent places to solve their housing problems in cities.However,finding an ideal housing place is a laborious task.Thus,to help migrants in this regard is urgently needed.As the key issue is to capture the causal relationship that users select houses,the task is thus a typical causal data fusion scenario.To address the task,we first extract users' and houses' features from multi-source urban data,including geographic data,traffic data,and e-commerce data.Thereafter,the features are fed to the proposed meta-learningbased deep neural network,named House Critic.In detail,in House Critic,user preference is used as the meta-knowledge to derive the parameter weights of house representations such that we can explicitly model the selection causality and accordingly value a given house.Finally,we conduct experiments on real-world datasets collected from Beijing,China.Experimental results demonstrate the advantages of the House Critic model over several baselines.Moreover,we have deployed a system by using the model in an e-commercial company to provide the housing suggestion service.? Aiming at investigating the fusion method for dynamic relationship-based data,we propose a novel deep learning-based policy.To further exploit the application of the proposed policy,we apply it based on a metro train scheduling application.Urban metros are foundational and have become the foremost public transit to modern cites,carrying millions of daily rides.Thus,shortening passengers' travel time for metros is urgently needed as the travel efficiency matters the work productivity of the city,bringing substantial economic benefits.For a train,each dwell decision has long-term impacts on the whole system.In addition,two complex factors affecting the dwell decision,i.e.,the status of passengers and the status of trains,are timeevolving.To capture the long-term impact and fuse the two factors,we propose a deep neural network,entitled Auto Dwell,as the policy.Specifically,to capture the long-term impacts,we propose to use a deep reinforcement learning framework to learn the network,which optimizes the long-term rewards of dwell time assigning in terms of passengers' waiting time on platforms and the journey time on trains.Moreover,Auto Dwell employs gated recurrent units and graph attention networks to capture the spatio-temporal correlations of the passenger flows,leverages attention mechanisms for capturing the interactions between the trains,and devise mechanisms to fuse the two aspects.Finally,extensive experiments on two real-world datasets collected from Beijing and Hangzhou,China,demonstrate the superior performance of the Auto Dwell model over several baselines,capable of saving passengers' overall travel time and in return directing metro operators to reduce train resources.The policy has been deploying on a metro operation company to assist in scheduling the dwell processes.