Commonly, normalization of ratings seems to mean adjusting values measured on different scales to a notionally common scale.

I have different metrics that measure different performance aspects of a server. I want to aggregate these values. The problem is:

  • The values are on different scale types (interval and ratio scale)

  • The values have different minimum and maximum values

    Most of them are between $0.0$ and $1.0$, where $0$ means complete absence of the measured attribute and $1$ means the attribute is maximized to the best characteristic it can gain.

  • One of the ratio scale values can be $[0, \infty]$. Theoretically, speaking, practically something around $500$. More is here better $0$ is absence, too.

  • The interval scale can be negative, positive and zero. Zero does not mean, absence. The less, the better. The more, the worse.

How do I approach this? I need to aggregate the results and feel a bit lost. I see two options here.

  1. Use a normalization, but which one? Ideally I would all map to $[0,1]$ so I can present them as percentage, although that would be unprecise for the one ratio and the one interval.

    I know min-max scaling (feature scaling?), but is that appropriate? What about z-score? With z-score however, it seems I will land again on a negative-positive scale.

  2. Use a tendency to aggregate, arithmetic mean (needs normalization), geometric mean (false semantic on zero results), adjusted geometric mean, etc.

How should I approach this, can someone advise me?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.