Comments posted to this topic are about the item The Shortcomings of Predictive Accuracy
One Orange Chip
I may be missing something, but you're talking about measuring accuracy in unsupervised learning techniques with clustering, especially K-Means clustering? Measuring accuracy in machine learning that is unsupervised learning with supervised learning is notoriously difficult and wrong. Though, I am not a data scientist bare in mind. I just know, especially with algorithms like K-Means, that finding K is the difficult part let along trying to prove it's accuracy using something like purity, which can be false positive based on the input of K.
xsevensinzx - Thursday, September 7, 2017 9:36 PM
To me, you're focused on a side issue here and not the main thrust of this article - which was to explore an alternate performance metric to the ubiquitous percentage accuracy. The idea was to showcase entropy as an alternate and to start to demonstrate how it might better capture the way that we (certainly I) would intuitively interrupt the results shown.
It is difficult to comment on the other parts of your comment. You draw a very black & white divide between clustering and classification that I don't agree with. In truth, these two methods are strongly complementary and share a strong degree of overlap. Keep in mind that machine learning methods are a tool that is used within a much larger process involving a lot of experimentation, repetition, iteration and evaluation.
One Orange Chip
I don't mean to make it so black and white. I agree they are very complementary.
Viewing 4 posts - 1 through 3 (of 3 total)