Adding decimals using the predicted probabilities

[nlassaux]

Hello there! First, thank you for maintaining the repo, it's been super helpful.

Since we are able to do ordinal regression, I've thought about getting artificial precision by leveraging the predicted probabilities. Even if we are working with 5 categories, I'd like to say that some datapoint is a 4.5 according to the probabilities we predict.

Example

Intuitively, I think it makes sense to say that [x1, x2, 0.75, x4] should to be closer to 4 than [x1, x2, 0.55, x4].

So, when x3=0.5, that'd be 3.0, and when x4=0.5, that'd be 4.0. And I'd expect something like [0.99, 0.95, 0.75, 0.25] to be returned as 3.5. When not on the edges, that would look like getting the decimals as computing x_i + x_i+1 - 0.5 for a predicted category of i.

Now, I'm not sure what to do for cases such as [0.25, 0, 0, 0] and [0.99, 0.95, 0.75, 0.80], on the edges...

Question

Would you say that such a modification makes any sense? Ever thought about doing something like that?