Exponential

Most fuzzy matching scores are linear - the penalty for being different by 1 unit is 1, by 2 units is 2 etc.

However, some penalties can be set to be exponential, meaning the difference is squared.

In these cases, a difference of 1 will still be 1, but a difference of 2 will have a score penalty of 4, a difference of 3 will have a penalty of 9, and so on.

Usage

Exponential penalties are ideal for removing outliers from matches.

They're often used when you have two or more of a particular attribute that has the same penalty - for example, you might have four different sensors measuring the height of a cake leaving your bakery oven.

We want to search for the cake with the perfect size - let's say 25% of the way up the tin. We've captured two examples of product runs in the captures below…

SensorCapture ACapture B
1130
2132
3130
410035

Capture A is obviously a terrible shape for a cake to be in. But mathematically, the average for Capture A is around 25%, while Capture B is almost 32%. Using linear penalties, the system would pick Capture A as a better fit than Capture B.

But using exponential penalties, Capture B is much better than Capture A.

Warnings

If you are mixing linear and exponential values, be aware of the possibility of a smaller, less significant scoring factor dominating one that is supposed to be more important.

We normally suggest giving exponential values small multipliers to reduce the chances of this happening.