1) What it is

  • Macro AUC = the average of AUC values computed per class.
  • Each class gets its own one-vs-rest AUC, then you average them.
  • Formula:

$\text{Macro AUC} = \frac{1}{K} \sum_{i=1}^K AUC_i$

where $K$ = number of classes, and $AUC_i$ = AUC of class $i$ vs all others.

2) Types of Macro AUC

  • Unweighted Macro AUC (most common):
    • Each class contributes equally, regardless of how many samples it has.
  • Weighted Macro AUC:
    • Each class AUC is weighted by class frequency.
    • Helps when classes are imbalanced.

3) Intuition

  • Macro AUC treats all classes as equally important.
  • Good if you care about rare classes just as much as frequent ones.
  • Contrast with Micro AUC:
    • Micro pools predictions first, so frequent classes dominate.

4) Example

Suppose a 3-class classification (A, B, C).

  • AUC(A vs rest) = 0.95
  • AUC(B vs rest) = 0.80
  • AUC(C vs rest) = 0.60

Macro AUC = (0.95 + 0.80 + 0.60) / 3 = 0.78

Even though class A is very good, the lower AUC for class C pulls the average down.

5) When to Use

  • Use Macro AUC if:
    • You want fair evaluation across all classes.
    • You care about minority classes.
  • Use Micro AUC if:
    • You want global performance, weighted by sample frequency.

Summary

  • Macro AUC = average of per-class one-vs-rest AUCs.
  • Treats all classes equally, regardless of class size.
  • Highlights poor performance on minority classes.
  • Contrast: Micro AUC pools predictions first, so frequent classes dominate.