I am trying to use a Neural Network to perform multiclass classification. The classes represent Insurance Risk Level. The most risky level is Level 1, the least risk corresponds to Level 10. The labels came from Unsupervised clustering unannotated insurance data. The architecture is:
Input Layer 43 features
Dense, 1000 neurons, ReLu activation
Dense 500 neurons, ReLu activation
Output 10 neurons, Softmax activation
In particular, for binary classification we can use single neuron as last layer. Then, we define the score to be the input value to this final layer. But for more than 2 classes I don't know what to do.
My question: Is there a way to obtain a Risk score that is consistent, that is, the greater the Level is, the smaller the score is ? I tried to use logits, not normalized probabilites, etc. But not sure if there is a way to get a single consistent score from logits ? . Thanks for any hint on this.