Research On Bank Real Estate Valuation Algorithm With Spatial Convolutional Network

The real estate mortgage loan is an important business of the bank.It needs to make a reasonable valuation of the true value of the real estate before issuing the loan,which is usually completed by a professional certified real estate appraiser entrusted by the bank.In recent years,the use of computer-aided real estate valuation technology is gradually emerging,but the accuracy of the existing valuation model still fails to meet the requirement.The purpose of this dissertation is to improve the existing real estate valuation algorithm from the perspective of spatial information.By mining spatial information from the satellite image without semantic annotation,and using the graph convolutional network to transmit the spatial information in a certain area,more sufficient spatial information can provided to real estate valuation and the purpose of improving the valuation accuracy can be achieved.With the help of these approachs,a more meaningful real estate valuation algorithm can be provided to the bank.Spatial information is an important factor of the real estate price.This dissertation proposes a method to mine large-scale spatial information by using spatial convolutional network.The main works of this research can be divided into two parts: firstly,a method is proposed to transform satellite image without semantic annotation into a low-dimensional dense vector related to the real estate price,namely SVCN(Spatial Vector Convolutional Network).Based on the principle of class activation mapping,we also propose a method called Deep-AOI(Deep Area of Interest)to get the attention information from the deep feature maps of the network.Thus,macro spatial information can be obtained and the visual information of satellite image can be used as neighborhood feature of regression model,which makes up for the limitation of traditional methods that only use micro POI(Point of Interest)spatial information.Secondly,under the consideration of spatial correlation,in order to transmit spatial information in a certain area and expand the receptive field of each sample,a method called SGCN(Spatial Graph Convolutional Network)is proposed.It contains two models: Field-GCN(Field Graph Convolutional Network)which can transmit spatial information without noise by dividing the features of node into public field and private field,controlling the information transmit between the nodes.In addition,under the consideration of spatial heterogeneity,we further improves the graph convolutional layer of Field-GCN,and proposes GW-GCN(Geographically Weighted Graph Convolutional Network).The network can adjust the weight of the feature transformation matrix adaptively according to the spatial position of the nodes,and conduct the geographically weighted regression on the graph convolutional network.The real data of second-hand real estate listing in Shenzhen is used for empirical test.The results show that the root mean square error of the traditional model can be reduced by 33.29% at most by introduing the spatial information from satellite image with SVCN,while the use of SGCN can further reduce the valuation error because they effectively aggregate the spatial information from adjacent real estates.The spatial convolutional network proposed in this dissertation not only provides new ideas and means for real estate valuation,but also can be used to study similar tasks with spatial correlation and spatial heterogeneity.