based on the sketch for the precision and recall computation and its link to the confusion matrix, we could cluster the classes by the network's error behaviour or underlying schema (e.g. animals supercluster, inanimate object supercluster) and compute the performance metrics inter and intra these clusters (as it's simply blocking of the confusion matrix).
based on the sketch for the precision and recall computation and its link to the confusion matrix, we could cluster the classes by the network's error behaviour or underlying schema (e.g. animals supercluster, inanimate object supercluster) and compute the performance metrics inter and intra these clusters (as it's simply blocking of the confusion matrix).
Could provide some interesting insights.
any thoughts?