Tuesday, 23 February 2016

POLYNOMIAL REGRESSION

Hello, welcome to my blog. I introduced the concept of linear regression in my previous posts by giving the basic intuition behind it and showing how it can be implemented in Python. In the last post, I gave a precaution to observe when applying linear regression to  a problem – Make sure the relationship between the dependent and independent variable is LINEAR i.e. it can be fitted with a straight line.

So, what do we do if a straight line cannot define the relationship between the two variables we are working with? Polynomial regression helps to solve this problem. 


Consider the scatterplot shown below


It’s obvious that the straight line in the scatter-plot does not effectively describe the trend in the data. If we were to use this line to predict the value of our dependent variable, we would get very poor results. Polynomial regression fixes this problem by adding a polynomial term to the equation for y – the predicted value. The equation now looks like this

                  y = a + b1*x + b2*x^2

This is a very simple polynomial equation, it is possible to have an equation with higher order polynomials. Here is the same scatter-plot shown above fitted with a second order polynomial. 



As you can see the line curved line generated by the second order polynomial fits the data points better than the straight line.

Going back to our case study of predicting house prices, let’s see how a second order polynomial does on the King County dataset.

MAKING POLYNOMIAL PREDICTIONS
First, let us visualize the fit of a second order polynomial on our data. I do this by using the regplot function from the seaborn library. This function automatically displays a best fit line for any scatterplot. In order to request a second order polynomial fit I pass the parameter order = 2 indicating I want to see the line generated by the second order polynomial that best fits my data. The scatterplot is shown below




Next, let us generate a model that uses a second order polynomial to predict house prices.
I am still using the ols function from the statsmodels.formula.api library in Python. The only difference is that I added I(sqft_living**2) which returns the values of sqft_living raised to power 2. 

Let’s view the results 




















Take note of the values I circled in red – they represent the parameter estimates of our model. We interpret them just like before. The first term ‘Intercept’ stands for a, ‘sqft_living’ is the coefficient of x i.e. b1 and ‘I(sqft_living ** 2)’ is the coefficient of x2 i.e. b2. Therefore, the equation for the second order polynomial that fits the data is


            y = 192500 + 73.4699*x + 0.0375*x^2


Predictions
On the test data set, this model predicts a price of about $374,000 for the first house. This is slightly worse than the prices predicted by the other models we have generated. The RSS (Residual Sum of squares – squared sum of the difference between the actual price and the price the model predicted) for this model on the test data is ≈ 2.54*1014.

Just to give some context, the RSS on test data for the multiple linear regression model was ≈ 2.45*1014. The RSS on test data for the simple linear regression model was ≈ 2.75*1014. So the polynomial regression model had the second best performance overall.

Summary
I introduced the concept of polynomial regression which helps us fit a non-linear relationship between two variables. I also implemented polynomial regression in Python and we used the model to make predictions.

Thanks for reading this post. I hope you enjoyed it. If you have any questions or comments feel free to leave a comment, I will be happy to answer any questions you have. Thank you once again for visiting my blog. Cheers!!!