Hi! First of all, thank you very much for this beautiful tool!
I was investigating how it works in detail while porting it to Python, and I wonder if it is an inaccuracy here in the calcPerClusterMappingMetric function, where you are finding the nearest reference centroid using cos distance, but the reference centroids have not been L2 normalised. The results are quite different if L2 normalisation is done (here using our Python implementation):
Hi! First of all, thank you very much for this beautiful tool! I was investigating how it works in detail while porting it to Python, and I wonder if it is an inaccuracy here in the
calcPerClusterMappingMetric
function, where you are finding the nearest reference centroid using cos distance, but the reference centroids have not been L2 normalised. The results are quite different if L2 normalisation is done (here using our Python implementation):