Study On Mass Appraisal Of Housing Price Based On Machine Learning ——A Case Study Of Chengdu Urban

In view of the fact that Chinese real estate tax has begun to pilot,but the actual value of housing is lack of objective evaluation basis.Thus,the study on high-precision mass appraisal of housing price has important real significance and practical value.Based on machine learning methods,taking the urban of Chengdu area as an example,the study of mass appraisal method for housing price was conducted.The main research contents and results of this thesis are listed as below:（1）In view of the current lack of basic feature data set of the standard mass appraisal about housing price,based on the prepocessing of housing information,points of interest（POI）,and road data set,the study about the construction of basic feature data about the urban housing price mass appraisal was carried out.After preprocessing of housing samples,POI,and roads data,the data set of housing price was constructed with the Arc GIS software.Based on the coordinate information of housing and POI,the features of distance and quantity related to POI were calculated.The feature selection based on machine learning was performed to obtain the basic feature data set which was related to the housing structure,community attributes and POI.（2）In view of the lack of information in feature extraction from POI and road data,a joint learning framework was proposed to learn the features of urban housing price according to POI and road data.Firstly,the convolutional neural network module was used to extract the features of kernel density maps about POI and road data sets.Then,the extracted features from the kernel density maps and the basic feature data set of urban housing price were input into a deep neural network,and the housing price was used as the label to train model to realize the learning of feature representation about housing price.The experiments were carried out to compare different structures and input sizes of joint learning frames,and the optimal joint learning frame structure was obtained with the residual network as feature extractor and the input size of 256×256.The experimental results showed that the urban housing price evaluation accuracies are improved compared with only using the basic feature data set or the kernel density map data set as input respectively.The bigger increase of the mean absolute error（MAE）for the model was up to 1 671.73￥/m~2,and the bigger increase of the relative mean absolute error（RMAE）was up to 9.15%.Thus,the proposed joint learning framework could be able to solve the problem of missing information,and it was effective in the mass appraisal of urban housing price.（3）In view of the lack of evaluation about the applicability of current methods for urban housing price mass appraisal,the study about the applicability of methods about urban housing price mass appraisal was conducted.Based on the data set of urban housing price constructed by feature extractor of residual network,the applicability of six mass appraisal methods of urban housing price based on machine learning was investigated.The experimental results showed that the accuracies of all the housing price mass appraisal methods had been improved on the data set from residual network.The accuracy of ridge regression improved more than other methods,and its RMAE was improved with the percent of 4.17%.However,extreme gradient boosting（XGBoost）and random forest（RF）had the best performance,with RMAE of 7.37%and 7.88%respectively.Error analysis about XGBoost,RF and deep neural network（DNN）,was also carried out.The experimental results showed that the error of RF was the least in Qingyang district,but XGBoost achieved the smallest errors in other districts.Therefore,some spational differences were lain in the study area,but the XGBoost yielded the best performance.（4）Due to spational differences of urban housing price evaluation methods,a study based on stacking ensemble learning was performed.Based on RF and XGBoost as the first learner and geographically weighted regression as the secondary learner,a double-layer mass appraisal approach of urban housing price was constructed.The experimental results showed that MAE,root mean square error（RMSE）,R~2 and RMAE of the method were up to 1 253.53￥/m~2,1 837.50￥/m~2,0.94,and 7.10%,respectively.Compared with XGBoost with the overall optimal learner,the MAE of stacking ensemble learning increased by 51.51￥/m~2,RMSE by 87.14￥/m~2,R~2 by 0.01,and RMAE decreased by0.27%.Therefore,the geo-weighted stacking ensemble learning method can improve the accuracy of mass appraisal for urban housing price.