Custom Classification Loss: which is the role of S?

Dear,

I have a classification problems with labels 1,..,5.

Since the label is a score related to a grade, I would to compute the Loss by computing the

distance between the triue label and the predicted label

So, if:

are the N points in my dataset,

is the TRUE label of , and if the predicted label is ,

is the weigh for point

is the cost of assigning the point in class to class

I would like to measure the loss ass:

However, how can I use the score to compute the Loss?

The score S contains ngative values. Whats the meaning of that score?

from the explanation in matlab help it seems that the more the S value is low (negative), the more the point is "distant from that class"

If I have 5 labels and for x(i) I have than this means that x(i) would have predicted label = 3

I used kfoldPredict to understand what's happening and it should be right.

Could anyone confirm me?

Answer 1

[Dinesh Yadav]

A classification score represents the confidence of the classifier into a class. The range of score depends on ensemble type. For example-

• · AdaBoostM1 scores range from -inf to inf.
• · Bag scores range from 0 to 1.

Therefore, if your scores range from -inf to +inf the class corresponding to the highest score is chosen. In the example you gave its label 3. The value 0 is highest among all.

I am providing the link to documentation below for score in kfoldPredict.

https://www.mathworks.com/help/stats/classificationpartitionedmodel.kfoldpredict.html

[Elena Casiraghi]

Thanks!!!