Linear Regression With Feature Scaling

Question

I have a general question about feature scaling in linear regression.

I have a dataset that is two years worth of data. The first year's worth of data for a specific column is completely different than the 2nd year's. I am assuming that maybe there were different attributes associated with calculating the 1st year's variable vs. the 2nd year.

Anyway, here is what the dataset looks like. I will show the first 6 rows of each year:

Date             Col1
2015-01-01       1500
2015-01-02       1432
2015-01-03       1234
2015-01-04       1324
2015-01-05       1532
2015-01-06       1424
.
.
.
2016-01-01         35
2016-01-02         31
2016-01-03         29
2016-01-04         19
2016-01-05         22
2016-01-06         32

When I want to forecast this dataset, obviously it is going to forecast results in the negative but in reality the data has just been rescaled in some way.

If I apply feature scaling as so, how do I revert back to my original dataset to make a forecast?

normalize <- function(x){
  return((x-min(x)) / (max(x)-min(x)))
}

scaled_data <- 
  df %>%
  group_by(Date %>%
  mutate(NORMALIZED = normalize(Col1))

You would need to save the min and max to be able to transform the result back to the original scale. — ekstroem, Jul 26 '17 at 20:20
Would the answer be then: `Predicted_Original_Scaled_Number = Predicted_Feature_Scaled_Number * (Max_Original - Min_Original) + Min_Original`? — nak5120, Jul 26 '17 at 20:26
Awesome thank you. Would you mind putting that as an answer to give you credit? — nak5120, Jul 26 '17 at 20:35

score 2 · Accepted Answer · answered Jul 26 '17 at 20:43

Sure. Might as well put it in a function although you did provide the answer yourself.

This one should be given the predicted value and the original vector

backtransform <- function(value, x) { value * (max(x) - min(x)) + min(x) }

or if you computed and kept the minimum and maximum then

backtransform2 <- function(value, min, max) { value * (max - min) + min }

Linear Regression With Feature Scaling

1 Answers1