Embedded Files
Here's how the metrics are computed, given a sample confusion matrix and the number of gold labels (aka support).
Confusion matrix
Confusion matrix
Key: N - Neutral, J - Joy, S - Sadness, A - Anger
PRED LABELS N J S A # of labelsT N 727 177 25 106 1035R J 209 252 11 33 505U S 57 8 47 9 121E A 54 19 8 60 141Per-class recall (r)
Per-class recall (r)
r(N) = 727/(727+177+25+106) = 0.702 = 70.2%......r(A) = 60/(52+19+8+60) = 0.436 = 43.6%Per-class precision (p)
Per-class precision (p)
p(N) = 727/(727+209+57+54) = 0.694 = 69.4%......p(A) = 60/(106+33+9+60) = 0.288= 28.8%Per-class F1-score (f1)
Per-class F1-score (f1)
f1(N) = 2[r(N)·p(N)]/[r(N)+p(N)] = 2[0.702·0.694]/[0.702+0.694] = 0.698 = 69.8%......f1(A) = ... = 0.344 = 34.4%Micro F1-score
Micro F1-score
micro-f1 = [1035 · 0.698 + ... + 141 · 0.344]/1802 = 0.603 = 60.3%Google Sites
Report abuse