TWM602260U - Real estate valuation system - Google Patents

Abstract

The present disclosure provides a real estate valuation system including a server and an electronic device coupled to the server. The server stores a valuation model, wherein the valuation model includes an ensemble learning model. The ensemble learning model includes multiple statistic models and the ensemble learning model receives input data and input the input data into the statistic models and generates a valuation value by a regression method. The server performs a validity verification to the ensemble learning model according to multiple refresh data in a time interval. When ensemble learning model does not pass the validity verification, the server refreshes multiple parameters corresponding to the ensemble learning model, and replace the ensemble learning model with a new ensemble learning model when the new ensemble learning model passes the validity verification. The electronic device performs an valuation operation using the valuation mode.

Description

Real estate valuation system

This disclosure relates to a real estate appraisal system, and particularly to a real estate appraisal system that automatically updates the appraisal model.

For the general public, the cost of buying a house is not expensive. Therefore, before the transaction is completed, several target items are often selected for condition, price confirmation and comparison. For example, the public can confirm the reasonableness of the price through various channels such as logging on the website at the actual price, real estate agents, and property owners quotations. However, this kind of work is time-consuming and laborious, and the inquiring object may also hide evil and promote good or overstate the price. In order to solve the above-mentioned housing purchase problem, an online real estate valuation system has been launched. However, most of the existing online real estate appraisal systems have a large gap between the appraisal results and the actual prices due to insufficient reference data. Therefore, how to provide a high-precision real estate appraisal system is a goal that those skilled in the art should strive for.

In view of this, the present disclosure provides a real estate valuation system that can automatically update the valuation model to provide more accurate valuation results.

This disclosure proposes a real estate valuation system, which includes a server and an electronic device coupled to the server. The server stores an evaluation model, where the evaluation model includes an integrated learning model. The integrated learning model includes a plurality of statistical models, and the integrated learning model receives input data and inputs the input data into the statistical model to generate estimated value in a regression manner. The server verifies the validity of the integrated learning model based on multiple updated data at intervals. When the integrated learning model fails the validity verification, the server updates multiple parameters corresponding to the integrated learning model, and replaces the integrated learning model with the new integrated learning model when the updated parameters pass the validity verification. The electronic device uses the valuation model to perform valuation operations.

Based on the above, the real estate valuation system of the present disclosure uses an integrated learning model as the valuation model. For example, the server receives updated data every month and inputs the updated data into the evaluation model for validity verification. If the evaluation model fails the validity verification, the server updates multiple parameters and performs validity verification again. If the re-modeled evaluation model passes the validity verification, the server can replace the original integrated learning model with the new integrated learning model, allowing the user to use the electronic device to perform the evaluation operation.

FIG. 1 is a block diagram of a real estate valuation system according to an embodiment of the present disclosure.

Please refer to FIG. 1, the real estate appraisal system 100 of an embodiment of the present disclosure includes a server 110 and an electronic device 120. The electronic device 120 is coupled to the server 110 through a wired or wireless network. The server 110 is, for example, a bank server and the electronic device 120 is, for example, a user's personal computer, notebook computer, smart phone, tablet computer, or other similar devices. The user can use the electronic device 120 to connect to the server 110 to perform real estate valuation operations.

In an embodiment, the server 110 stores an evaluation model, where the evaluation model includes an integrated learning model. The integrated learning model includes a plurality of statistical models, and the integrated learning model receives input data and inputs the input data into the statistical model to generate estimated value in a regression manner. In other words, the output of multiple statistical models in the integrated learning model can form a stacking model of Linear Regression. Input data can include credit information, real price registration data, point of interest data (POI), economic index data (for example, it can include a variety of economic indexes), mortgage index compiled by house price data, and government potential Regional official data, bank collateral data and other real estate-related data, and the output data is, for example, real estate valuation results. The input data can also include credit collateral data, official data of a government point of interest, and housing price data (for example, it contains data from multiple builders and the housing price data is not limited to those provided by the builder. It can also include the housing price data of multiple builders. Individual mortgage index), potential area data (can contain a variety of related data at the same time), and vacant home index data (can contain multiple related data of this type at the same time). The server 110 verifies the validity of the integrated learning model according to multiple updated data at time intervals (for example, every month). When the integrated learning model fails the validity verification, the server 110 updates multiple parameters corresponding to the integrated learning model, and replaces the integrated learning model with the new integrated learning model when the new integrated learning model after the updated parameters passes the validity verification. For example, the server 110 deploys the evaluation model to the application host to provide evaluation operations, and updates the model file of the evaluation model Application Program Interface (API) for the front desk to call the latest evaluation result. The public uses the electronic device 120 to use the valuation model to perform valuation operations.

In an embodiment, the server 110 may establish an evaluation model according to two stages. The first stage can include using a variety of statistical models with different advantages and referring to a variety of characteristic parameters that affect prices, using a certain price parameter source as the base (without limiting the total number of parameters), and adjusting the sampling content based on externally input object information. Automate the use of sample repeated sampling (that is, take and put it back) to generate multiple subsets (Subsets), distribute the input to multiple statistical models established, and use regression methods (without limitation on algorithm types) to obtain various types The statistical model estimates the predicted mean. In the second stage, another model can be designed to use regression methods (without limitation of algorithm types), and the above-mentioned various statistical models can be automatically integrated and trained and the advantages of various models can be maintained. For example, for example, some models can have different weights for multiple predicted values, and some models can optimize the weights of classification data based on the previous model to increase the error to enhance the model, and some models have higher linear independence. In this way, by integrating multiple types of models, higher accuracy of valuation and forecasting than a single type model can be generated.

In one embodiment, the input data may include real price registration, potential area, housing price index, economic index, vacant house index, POI data, etc. The real price registration is the real property registration data. Potential areas include soil liquefaction potential and flooding potential. The house price index includes, for example, the Xinyi House Price Index and the Cathay Pacific Real Estate Index. Economic indicators include the average recurring salary of the most recent month/quarter/year, consumer price index, prosperity light score, unemployment rate, home loan interest rate, stock price index, leading indicators of prosperity, industrial production index, total money supply count , Average loan amount, gross domestic production and savings rate, etc. The index of vacant housing can include low electricity consumption data and new surplus housing data. POI data can include people's livelihood consumption, educational institutions, public transportation, financial services, life and leisure, medical institutions, aversion facilities, etc. People's livelihood consumption may include financial card special stores, convenience stores, wireless hotspots, Internet cafes, movie theaters, parking lots, markets, etc. Educational institutions can include middle and high schools, colleges, primary schools, preschool education, libraries, nursery centers, tuition schools, etc. Public transportation can include Taiwan Railway Station, bicycle rental station, bus stop sign, high-speed rail station, MRT station, etc. Financial services may include banks, automated teller machines, etc. Life and leisure can include general hotels, international tourist hotels, and leisure parks. Medical institutions may include medical units and pharmacies. Suspected facilities can include gas stations, industrial parks, funeral facilities, religious sites, gas stations, garbage dumps, temples and altars, incinerators, power plants, substations, social housing, fire stations, fault information areas, etc.

In one embodiment, the actual price registration information may include township, city, county, latitude and longitude, number of land transfer pings, land use zoning, house age, number of transactions, transfer level, total number of floors, building form, and main purpose , Main building materials, applicable building regulations time, number of building transfers, current structure of the building, whether there is a management organization, unit price per ping, parking space category, number of parking spaces transferred, whether parking space prices are split, etc.

In an embodiment, the server 110 may update multiple parameters corresponding to the valuation model according to a grid search algorithm or a genetic algorithm. Take the grid search algorithm as an example. The grid search algorithm can try every possibility in a loop among all candidate parameter combinations, and use the best performing parameter combination as the final result. In one embodiment, both the random forest model and the eXtreme Gradient Boosting (XGBoost) model can use a grid search algorithm to automatically find the best parameter combination of the evaluation model. Table 1 shows the model, parameters and parameter descriptions of the grid search.

Table I model parameter Parameter Description Random forest model n_estimators Number of trees in random forest max_depth Maximum depth of trees in random forest min_samples_split Minimum number of samples required for branch min_samples_leaf Lower limit of the number of samples of leaf nodes Extreme gradient boosting model eta Learning rate max_depth Maximum depth of tree subsample Sample sampling ratio colsample_bytree Variable sampling ratio gamma Lower limit of loss function value min_child_weight Sum of sample weights of child nodes

On the other hand, the server 110 may also update multiple parameters of the corresponding valuation model according to the genetic algorithm. The efficiency of using genetic algorithms to determine the most suitable parameters is much greater than grid search.

In one embodiment, the updated data includes a plurality of real estate data. When the difference between the output result of the real estate data input to the valuation model and the plurality of predetermined valuation results corresponding to the real property data is less than or equal to the threshold value, the output result passes the validity verification . When the difference between the output result and the predetermined evaluation result is greater than the threshold value, the output result does not pass the validity verification.

For example, the server 110 may calculate the Mean Absolute Percentage Error (MAPE) and hit rate corresponding to the output result. When the average absolute error percentage is less than or equal to a predetermined percentage (for example, 15%) and the error is within the first percentage (for example, 10%), the hit rate is greater than or equal to the second predetermined percentage (for example, 50%) and the error is within the second If the hit rate in the percentage (for example, 20%) is greater than or equal to the third predetermined percentage (for example, 80%), the output result passes the validity verification. Through double verification of average absolute error percentage and hit rate, the accuracy of validity verification can be greatly increased.

，其中y為估價結果實際值且ŷ為估價結果預測值且n為樣本數。命中率是指在特定誤差範圍內，預測值落於該區間的機率，命中率越高，表示預測值接近實際值的機率越高。命中率的計算公式為

，

，其中y為估價結果實際值且ŷ為估價結果預測值且

為信心水準且N為測試樣本數且n為命中區間樣本數。 The average absolute error percentage is used to measure the ratio of the difference between the actual value and the predicted value to the actual value. The error is taken as the absolute value, and the magnitude of overestimation or underestimation is equal, and the errors will not offset each other. The formula for calculating the average absolute error percentage is

, Where y is the actual value of the valuation result and ŷ is the predicted value of the valuation result and n is the number of samples. The hit rate refers to the probability that the predicted value falls within the interval within a certain error range. The higher the hit rate, the higher the probability that the predicted value is close to the actual value. The formula for calculating the hit rate is

,

, Where y is the actual value of the valuation result and ŷ is the predicted value of the valuation result and

Is the confidence level and N is the number of test samples and n is the number of samples in the hit interval.

In one embodiment, the server 110 can create multiple statistical models, which can effectively improve accuracy compared with a single model. The server 110 uses an integrated learning (Ensemble Learning) technology to use a combination of three types of models: Bootstrap aggregating (also known as Bagging), Boosting, and Stacking as a predictive model base for real estate valuation. The calculation method can be divided into two stages. The first stage mainly uses the random forest method derived from self-service aggregation models with different characteristics and the extreme gradient boosting method (eXtreme Gradient Boosting, XGBoosting) derived from the boosting model to output multiple sub-model prediction values. In the second stage, the stacking model is used to integrate the results of the two models in the first stage to preserve the advantages of the original first stage models and eliminate the shortcomings of the single model, further improve the prediction accuracy and output the final prediction results. The details of the above related models are shown in Table 1 below.

Table I Two-stage operation The first stage second stage Algorithm Random forest method (self-service aggregation + decision tree) Limit gradient boosting method (boosting + decision tree + gradient descent method) Stacking method Model architecture Self-service aggregation (strong model, sub-model samples are retrieved and redrawn, with high repeatability and easy interference) Promotion (weak model, sub-models separate incorrect data in a progressive manner) Stacking method concept Repeat sampling from the same data source, and add the predicted values of the sub-models for each sampling Use multiple models to sequentially separate error data and weight, reduce deviation, and finally retain high-precision data Use linear regression, support vector regression or decision tree to integrate the results of the two types of models in the first stage of training advantage Low variability, low collinearity, can detect the validity of each variable for deletion, can detect the optimal number of variables and sub-models, and have rapid calculation capabilities Low deviation value, less variable explanatory problems, rapid calculation ability (regularization reduces model complexity) Provide prediction results combining the advantages of various models in the first stage Disadvantage High deviation High degree of change, more noise N/A Data verification k-fold cross validation

In one embodiment, the server 110 can be modeled in 7 major dimensions, including: real price registration, types of surrounding facilities (POI data), housing price index, potential area data, economic indicators, vacant housing indicators, and local levy Approval system (e-Loan system) collateral information and regional classification (loan number) data. Factors including multiple parameters, historical data and expected future economy are taken into consideration.

In the "net registration" aspect, based on transaction, target, and price information, there are more than 1.3 million real price registration data, and a comprehensive search for similar objects within 500 meters/1000 meters in diameter.

In the "POI data" dimension, the facility data of the government open information platform (e.g. https://data.gov.tw/) is connected in 7 categories (e.g. people’s livelihood consumption, educational institutions, public transportation, life and leisure, Financial services, medical institutions, disgusting facilities), conduct a comprehensive search of the surrounding area to increase and decrease valuations.

In the "House Price Index" aspect, according to the database of Xinyi Housing and Cathay Pacific Construction, according to the transaction volume, price, and sales rate, the trend of housing prices in various regions, counties and cities is known, and the valuation increases and decreases according to the district.

In the "potential area data" aspect, the estimation of soil liquefaction, flooding degree increase and decrease is obtained according to the government's public information platform (for example, https://data.gov.tw/).

In the "Economic Indicators" dimension, according to the overall statistical database (for example, http://statdb.dgbas.gov.tw/pxweb/Dialog/statfile9L.asp) 12 kinds of economic indicators (for example, the consumer price index) , Prosperity indicator score, unemployment rate, leading indicators of prosperity, industrial production index, money supply, total currency count, stock market stock price index, gross domestic production, savings rate, average salary of loan lines, interest rate of home loans), combined with economic trends Change and adjust model valuation to improve accuracy.

In the "Vacancy Index" dimension, the number of households with low electricity consumption is used to build vacant housing based on the real estate information platform of the Ministry of the Interior (e.g. https://pip.moi.gov.tw/V2/E/SCRE0104.aspx) The information determines the supply and demand of surrounding houses.

In the "Collateral Data and Regional Classification Data of the Bank's Collection and Approval System", we have obtained about 100,000 or more loan appraisal data based on the Bank's collection and review system (e.g. e-Loan system) and selected nearby objects and the The area corresponds to the loanable percentage, and the regional grade data refers to the regional loanable percentage.

In summary, the real estate appraisal system disclosed in the present disclosure uses an integrated learning model as the appraisal model. For example, the server receives updated data every month and inputs the updated data into the evaluation model for validity verification. If the evaluation model fails the validity verification, the server updates multiple parameters and performs validity verification again. If the re-modeled evaluation model passes the validity verification, the server can replace the original integrated learning model with the new integrated learning model, allowing the user to use the electronic device to perform the evaluation operation.

Although this disclosure has been disclosed in the above embodiments, it is not intended to limit the disclosure. Anyone with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of this disclosure. Therefore, The scope of protection of this disclosure shall be subject to those defined by the attached patent scope.

100: Real estate valuation system 110: server 120: electronic device

FIG. 1 is a block diagram of a real estate valuation system according to an embodiment of the present disclosure.

100: Real estate valuation system

110: server

120: electronic device

Claims (7)

A real estate valuation system including: A server; and An electronic device coupled to the server, wherein The server stores an evaluation model, wherein the evaluation model includes an integrated learning model, the integrated learning model includes a plurality of statistical models, and the integrated learning model receives an input data and inputs the input data into the statistical models and uses a The return method produces an estimated value; The server performs a validity verification on the integrated learning model according to a plurality of updated data at a time interval; When the ensemble learning model fails the validity verification, the server updates multiple parameters corresponding to the ensemble learning model, and when a new ensemble learning model after updating the parameters passes the validity verification, the new The integrated learning model replaces the integrated learning model; and The electronic device uses the evaluation model to perform an evaluation operation. The real estate appraisal system according to claim 1, wherein the server calculates an average absolute error percentage and a hit rate, and judges whether the integrated learning model passes the validity verification based on an average absolute error percentage and/or a hit rate . The real estate appraisal system described in claim 2, wherein the average absolute error percentage is

, Where y is the actual value of the valuation result, ŷ is the predicted value of the valuation result and n is the number of samples, the hit rate is

,

, Where y is the actual value of the valuation result and ŷ is the predicted value of the valuation result and

Is the confidence level and N is the number of test samples and n is the number of samples in the hit interval. The real estate appraisal system according to claim 1, wherein the server updates the parameters corresponding to the integrated learning model according to a grid search algorithm or a genetic algorithm. The real estate valuation system according to claim 1, wherein the statistical models include a random forest model. The real estate valuation system according to claim 1, wherein the statistical models include an extreme gradient boosting (eXtreme Gradient Boosting, XGBoost) model. Such as the real estate appraisal system described in claim 1, wherein the input data includes a credit collateral data, a real price registration data, an official government point of interest data, an economic index data, a housing price data, a potential area data, Index data of an empty house.

Images