begin the measure functions with measure?

[PhilippPro]

For example measureACC or better ACC or better accuracy?

[giuseppec]

Not sure about the naming. We probably want to avoid name clashes. Maybe ACC uppercase is a good choice?

[jakob-r]

I prefer a unified prefix which makes it easier to go through measures and this also avoids name clashes with any other package.

[PhilippPro]

But if we begin with "measure" we have the name clash with mlr.

[jakob-r]

Some ideas:

• meAUC, meACC, meRMSE
• mesAUC, mesACC, mesRMSE
• msAUC, msACC, msRMSE
• perfAUC, perfACC, perfRMSE

or err for everything that we minimize and perf for everything we maximize?

[giuseppec]

I don't like errRMSE because it then stands for error Root Mean Square Error. But I think I like the idea to have a naming that already makes a clear distinction between measures that should be minimized and measures that should be maximized. Maybe loss and gain are more appropiate then err and perf as loss function / gain function are well-established terms (and both terms have the same length, i.e. 4 characters)?

• MMCEloss and ACCgain
• lossMMCE and gainACC

[mllg]

You could also put all measures in a dictionary-like object (like we do in mlrng) and implement a function

measure(y, yhat, [measure])

with [measure] being either a string or an actual (parametrized) measure.

Pros:

• You can list all measures (measures$list())
• You can query measures w.r.t. their properties (measures$list(properties = "multiclass")
• No name clashes

Cons:

• Users have to manually do something like mmce = measures$get("mmce") to get the measure
• Documentation is slightly more difficult

[jakob-r]

Can't we do mmce = measures$get("mmce") automatically for the most common ones?

[mllg]

Sure, that would work. But the problem with the name clashes persists.