Scaling values
Today while reading Data Science Algorithms in a Week, from packt, I came across the concept of rescaling values so that when measuring their distances they would be more relevant. The dataset consisted of "House Ownership":
| Age | Annual income in USD | House ownership status |
|---|---|---|
| 23 | 50,000 | Non-owner |
| 37 | 34,000 | Non-owner |
| 20 | 100,000 | Owner |
| 35 | 130,000 | Owner |
etc..
The aim being to predict whether a person that is 50 years old with an income of $80,000, would own a home so that he could be targeted for home insurance.
k-Nearest Neighbor's are currently being covered and for this exercise a k = 1 is to be used.
Using either a Euclidean or Manhattan distance wouldn't matter because the distances between these values are too great. In comes rescaling!
The formula use is:
So for this dataset, both the Age and the Annual income wound have to be adjusted.
After scaling, the adjusted dataset would look something like this:
| Age | Scaled age | Annual income in USD | Scaled annual income | House ownership status |
|---|---|---|---|---|
| 23 | 09375 | 50,000 | 2 | Non-owner |
| 37 | 53125 | 34,000 | 4 | Non-owner |
| 20 | 0 | 100,000 | 7 | Owner |
| 35 | 46875 | 130,000 | 1 | Owner |
| 50 | 9375 | 80,000 | 5 | ? |
Now a 1-NN algorithm with a Euclidean metric could easily be used to find out if the last person is more than likely to own a home. Without the rescaling, the algorithm would have yielded different results.
Keeping it short today, but hopefully it was a helpful tip!