I am applying **Polynomial Regression** to my data, however the parameters theta were always **=0**,

i noticed that my y data or output is too large in the order of 100000 so i **normalized y**, i got very good predictions.

However what can i do if i need to get the predictions based on the **real y values**, meaning how can i **denormalize** ? wont the parameters be wrong afterwards somehow since i got them off of the normalized data ?

Note: i also normalize my input x data

**Contents**hide

#### Best Answer

Most techniques for "normalization" are invertible. If you think of your normalization procedure as a function $f$, this means we can find a function $f^{-1}$ such that $y = f^{-1}(f(y))$. This means you can apply $f^{-1}$ to your predicted value to get back to the original "unnormalized" scale, or denormalize the output using your parlance. Here's some common ones.

Standard score $$ begin{align*} f(y) &= frac{y – mu}{sigma}\ f^{-1}(y) &= sigma y + mu end{align*} $$ Scaling [0,1] $$ begin{align*} f(y) &= frac{y – y_min}{y_max – y_min}\ f^{-1}(y) &= y (y_max – y_min) + y_min end{align*} $$

### Similar Posts:

- Solved – Motivating sigmoid output units in neural networks starting with unnormalized log probabilities linear in $z=w^Th+b$ and $phi(z)$
- Solved – Motivating sigmoid output units in neural networks starting with unnormalized log probabilities linear in $z=w^Th+b$ and $phi(z)$
- Solved – Posterior distribution of Normal Normal-inverse-Gamma Conjugacy
- Solved – Conditional Expectation of Log-Normal Distribution
- Solved – Conditional Expectation of Log-Normal Distribution