cleanzr / clevr Goto Github PK

Suppose true_pairs is represented as an integer array with canonical representation

      [,1]  [,2] 
 [1,]   25   107
 [2,]  106   175
 [3,]  108   203

but pred_pairs is represented as a character array with canonical representation

      [,1]  [,2] 
 [1,] "106" "175"
 [2,] "107" "25" 
 [3,] "108" "203"

The canonical representation of the pairs is different, due to the different lexicographic ordering of strings vs integers.
This case should raise a warning (at a minimum) in contingency_table_pairs. The default behavior should potentially be to coerce both pairs to character mode?

Matt Barnes review several measures for evaluating ER in a tech report. We have already implemented most of them, but could consider adding:

• cluster precision, recall, F-score
• closest cluster precision, recall, F-score
• purity and K measure
• generalized merge distance