In this lab, you'll be able to validate your model using train-test-split.
You will be able to:
- Calculate the mean squared error (MSE) as a measure of predictive performance
- Validate the model using the test data
This time, let's only include the variables that were previously selected using recursive feature elimination. We included the code to preprocess below.
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
%matplotlib inline
from sklearn.datasets import load_boston
boston = load_boston()
boston_features = pd.DataFrame(boston.data, columns = boston.feature_names)
b = boston_features["B"]
logdis = np.log(boston_features["DIS"])
loglstat = np.log(boston_features["LSTAT"])
# minmax scaling
boston_features["B"] = (b-min(b))/(max(b)-min(b))
boston_features["DIS"] = (logdis-min(logdis))/(max(logdis)-min(logdis))
#standardization
boston_features["LSTAT"] = (loglstat-np.mean(loglstat))/np.sqrt(np.var(loglstat))
X = boston_features[['CHAS', 'RM', 'DIS', 'B', 'LSTAT']]
y = None
A good way to compare overall performance is to compare the mean squarred error for the predicted values on the train and test sets.
If your test error is substantially worse then our train error, this is a sign that our model doesn't generalize well to future cases.
One simple way to demonstrate overfitting and underfitting is to alter the size of our train test split. By default, scikit learn's built in method allocates 25% of the data to the test set and 75% to the training set. Fitting a model on only 10% of the data is apt to lead to underfitting, while training a model on 99% of the data is apt to lead to overfitting.
Iterate over a range of train-test split sizes from .5 to .95. For each of these, generate a new train/test split sample. Fit a model to the training sample and calculate both the training error and the test error (mse) for each of these splits. Plot these two curves (train error vs. training size and test error vs. training size) on a graph.
<matplotlib.legend.Legend at 0x1a24d6cef0>
Repeat the previous example, but for each train-test split size, generate 100 iterations of models/errors and save the average train/test error. This will help account for any particularly good/bad models that might have resulted from poor/good splits in the data.
<matplotlib.legend.Legend at 0x1a26e93438>
What's happening here? evaluate your result!
Congratulations! You now practiced your knowledge on MSE and on using train-test-split.