class-1 represents 0.01, class-i represents 0.01*i, class-100 represents 1.00.
Thus, when the classifier predicts the class-y and it should have predicted class-(y+1) there is a small error so we can accept class-y.
Is there a way to express this behaviour in a neural network? Maybe with a distribution or something?
PS: Not interested in regression.
-
$\begingroup$ Hi, welcome to Data Science StackExchange! Although you are not interested in regression, this is typically a regression problem. Is the original problem more complicated than what you express here? If yes, please provide additional details. $\endgroup$ – Romain Reboulleau Nov 11 at 12:26
-
$\begingroup$ I am new to Machine Learning and I like experimenting. Lately I had the curiosity for what I asked above and could not google it the correct way I guess to find a proper answer. What I was looking for was Ordinal Categorical Classification, given by @serali . $\endgroup$ – E. Vasilopoulos Nov 11 at 19:36
-
$\begingroup$ Cool, I learned something! Feel free to upvote the answer and mark it as accepted. $\endgroup$ – Romain Reboulleau Nov 11 at 20:36
-
$\begingroup$ Ok, I will (not allowed to upvote yet btw). Thank you. $\endgroup$ – E. Vasilopoulos Nov 11 at 23:21
Correct me if I am wrong but If I understand your question correctly, what you want is a classifier such that classes close to each other (say class 2 and class 3) are preferable to those far away (class 2 and class 99). If this is the case, this problem is called "Ordinal Categorical Classification".
I was working on a similar problem a while ago, and I found this loss function during my research. I ended up not using it so I don't really know how good it works but anyways, hope that helps.
-
$\begingroup$ Yes this is exactly the answer I was looking for. I did not know about Ordinal Categorical Classification. Thank you very very much. $\endgroup$ – E. Vasilopoulos Nov 11 at 19:31