Azure 机器学习与 Python 的过拟合/欠拟合机器学习模型

Question

I'm learning how to perform Machine Learning with Azure ML Studio. At the moment, I've only played around with Machine Learning using Python.

I have run identical Machine Learning projects using both Azure ML and Python to see how close the results of each product with the Root Mean Squared Errors (RMSE). So far the RMSE has been widely different for Azure ML and Python.

I can't figure out why the RMSE is so far apart. The only reason I can think of is because of the way Python 'fits' the model on the training data. Python uses the following code to fit the training data

lr = LinearRegression(labelCol='xxxx')
lrModel = lr.fit(train_data)

However, I don't know how Azure ML fits the training data.

Can someone let me know how Azure ML accomplishes fitting the training data?

Answer 1

Im guessing you probably used RMSE = √( 1/n ∑ (y_i - pred_i)^2 ) to calculate the RMSE in python, where y are the true labels, pred are the predicted ones in the linear regression?

I can imagine Azure using a slightly different term, namely RSE = √( 1/(n-2) ∑ (y_i - pred_i)^2 ), in which the Bessel's Correcting term 1/(n-2) is used instead. This is used to correct for the bias of fitting 2 parameters (assuming the linear regression is only fitting the slope and the intercept, otherwise 1/(nk) is the Bessel's Correction when fitting k parameters in a multivariate linear regression)

Try it out! I cannot explain however why the difference between python and Azure is so large, since the Bessel's Correcting term should only make a small difference.