Design And Implementation Of Real Estate Price Forecasting System Based On Multimodal Information Fusion

With the rapid development of the national economy,real estate as a pillar industry has played an increasingly important role in promoting the economy,and real estate price forecasts are taking the lead in real estate investment and second-hand housing transactions.The biggest challenge in housing price forecasts is that real estate prices are affected by many factors.At present,the mainstream price forecasting method is to establish a mathematical model.However,due to the nonlinear relationship between real estate prices and influencing factors,and the influencing factors are interrelated,the prediction accuracy still has room for improvement.Based on the above problems,this thesis designs a real estate price forecasting system based on multimodal information fusion,aiming to use time series information and house image information to improve prediction accuracy.For time series information,this thesis proposes the Smoothing Based Autoregressive Integrated Moving Average Model to solve the problem of high differential rank when use the traditional ARIMA on real estate price forecast.For house image information,this thesis proposes the Fine-tune Based Embedding Generation Algorithm.The algorithm extracts the house image feature,and uses Transfer Learning to transfer the house image classification ability to embedding generation task which can avoid labeling a large amount of house data.This thesis proposes a Multimodal Deep Learning Model to establish cross-modal information fusion by using the Shared Representation Layer.The fusion makes the model have stronger generalization ability and higher precision than the single mode model.This thesis designs and implements the real estate price forecasting system based on imultimodal information fusion.The system includes four modules:front-end interaction,algorithm capability,data management and system management.It supports access through the website and browser plug-in.The model training is offline.The offline model parameter file is maintained by the data management module.When the user submits the query request,the algorithm capability module performs online calculation and returns to the fr-ont-end interaction module to display the result,and uses system management module for the management and control of the entire system service,interface,and operation log.Through comparison experiments and system tests,the real estate price forecasting model based on multimodal information fusion proposed in this thesis has better performance on Mean Squared Error than the comparison model,and the system function is complete,with high availability and robustness.