Research On Valuation Of Assets Based On Machine Learning

With China's reform and opening up,the economy develops rapidly.As one of the most important assets,real estate also grows rapidly.Since China is going to take a tax reform,the real estate tax may be implemented in the future.It is significant to achieve a quick,correct,automatic real estate valuation.The study references a large number of papers and the latest research results,and proposes a new method based on machine learning to solve some outstanding problems in real estate valuation system,such as insufficient mining unstructured data and poor fitting of regression models.The new real estate valuation system has two parts,includes the floor plan classification based on transfer learning and house price valuation based on ensemble learning.Researches in this paper includes:(1)Research on extracting features from the floor plan based on transfer learning.As the floor plan hides a large amount of high-value information,the paper proposes a way to extract features based on transfer learning.It uses segmentation algorithm to split wall lines from the pre-processing floor plan,and trains VGGNet-16 model from ImageNet data sets by transfer learning to extract image features.(2)Research on dimensionality reduction based on self-encoder and house classification based on Multi-layer Perceptron.The dimension of data is so large that it is difficult to be processed.The paper reduces the dimension of data from 25088 to 500 by self-encoder,then trains a Multi-layer Perceptron model to divide house into 6 different types.The accuracy of classification is around 0.894.(3)Research on housing price valuation based on ensemble learning.The traditional regression models are difficult to improve the accuracy.However,ensemble learning can mix different models together to increase result by adjusting their weights.The paper proposes a house price valuation model based on ensemble learning.The new model connect with four different models,includes Ridge Regression,Random Forest,Factorization Machine and XGBoost.The final result reaches around 0.0959 in RMSE(Root Mean Square Error).