Neural network method -- Boston house price (regression problem)

One 、 The return question —— Linear regression and softmax Return to

Linear regression — It refers to a kind of method to model the relationship between one or more independent variables . In the field of natural science and Social Science , Regression often represents the relationship between input and output .

Two 、 Boston house price ( The return question )

Problem description ：

        The dataset is a regression problem . The number of observations for each class is equal , share 506 Observations ,13 Input variables and 1 Output variables .
　　 Each data contains detailed information about the house and its surroundings . Including the urban crime rate , Nitric oxide concentration , The average number of rooms in the house , The weighted distance to the central area and the average house price from the house, etc .

• CRIM： Crime rate per capita in cities and towns .
• ZN： More than 25000 sq.ft. The proportion of .
• INDUS： The proportion of Urban Non retail commercial land .
• CHAS： Charles River variables （ If the boundary is a river , Then for 1; Otherwise 0）.
• NOX： Nitric oxide concentration .
• RM： The average number of rooms in the house .
• AGE：1940 Proportion of self use houses built before .
• DIS： Weighted distance to Boston's five central areas .
• RAD： Proximity index of radial road .
• TAX： Every time 10000 The full value property tax rate of US dollars .
• PTRATIO： The proportion of teachers and students in the city .
• B：1000（Bk-0.63）^ 2, among Bk The proportion of black people in the town .
• LSTAT： The proportion of the population in the lower ranks .
• MEDV： The average price of a house , In thousands of dollars .

1. First read in the data , The code is as follows ：

#1. Data read in 
from sklearn.datasets import load_boston
import numpy as np
import pandas as pd
data=pd.read_csv('C:/Users/bby/ machine learning /boston_house_prices.csv')# Read in file 
data.head()

2. Divide the data into test data set and training data set

from sklearn.model_selection import train_test_split
X=data.drop('MEDV',axis=1)   # Generate feature set 
y=data['MEDV']               # Generate labels Set 
X=X.values                   # Convert to array
y=y.values                   # Convert to array
X_train,X_test,y_train,y_test=train_test_split(X,y,random_state=33,test_size=0.25)

3. Input the data with a wide range of values into the neural network , There's a problem . The network may automatically adapt to this data with different value ranges , But learning is bound to become more difficult . So standardize the data .

from sklearn.preprocessing import StandardScaler
ss_X=StandardScaler()
scaler_X=ss_X.fit(X_train)
X_train=scaler_X.transform(X_train)
X_test=scaler_X.transform(X_test)

4. Build a model , Choose the appropriate optimization method , The number of hidden layers can be adjusted .

from sklearn.neural_network import MLPRegressor# Full connection 
model = MLPRegressor(solver='lbfgs', hidden_layer_sizes=(15,15), random_state=1)
model.fit(X_train, y_train)

5. Model to evaluate , The most commonly used loss function for regression problems is the mean square error MSE(mean  aquared error) The code is as follows ：

from sklearn.metrics import r2_score,mean_squared_error,mean_absolute_error
print(' Training set regression evaluation index ：')
model_score1=model.score(X_train,y_train)  
print('The accuracy of train data is',model_score1)  # Goodness of fit value 
print(' Test set regression evaluation index ：')
model_score2=model.score(X_test,y_test)  
print('The accuracy of test data is',model_score2)  
y_test_predict=model.predict(X_test)
mse=mean_squared_error(y_test,y_test_predict)   # Mean square error 
print('The value of mean_squared_error:',mse)
mae=mean_absolute_error(y_test,y_test_predict)  # Mean absolute error 
print('The value of mean_absolute_error:',mae)

Try super big data .