Timeline for What is the relationship between mean squared error and classification error?

6 events

when toggle format	what		by	license	comment
Apr 19, 2020 at 21:44	history	bounty ended	gator
Apr 18, 2020 at 15:55	comment	added	EdM		@gator there are many ways that MSE and CErr as criteria can lead to different models. If you have pre-specified the cost tradeoff between false-negative and false-positive classifications you can consider a wide range of proper scoring rules to minimize expected cost in your application. This thread and this answer provide more detail and links to further reading.
Apr 18, 2020 at 0:36	vote	accept	gator
Apr 18, 2020 at 0:35	comment	added	gator		To make what I said short, my neural outputs in the CErr optimized network might be within the range `[0.51, 1.00]` whereas the MSE optimized network might be `[0.90, 1.00]` to give an example.
Apr 18, 2020 at 0:34	comment	added	gator		Interesting. My network does consider, in the case of a binary classification problem, a output of `0.51` for one class as belonging to that class. I guess even if CErr is optimized, CErr is not a suitable measure because it doesn't show confidence in classification. If an CErr optimized network only cared about cutoffs of probabilities, an MSE optimized network might not and my neural outputs might be closer to `1.00` than the CErr optimized network, thus being more confident in classification.
Apr 17, 2020 at 23:27	history	answered	EdM	CC BY-SA 4.0