Hello,
the AI-model is a simple classification of audio data:
8 classes, with 8 different records.
The common way to going on is to test an audio file – that fits (with prediction/accuracy) to each classes, and one class will fit as best. This is the procedure.
Now, i must add a feature, because this simple procedure will not solve my problem
I have very important informations about this classes: 7 classes(sounds) are “false”, and one ist “right”; and the 8 sounds are equidistant to the next, such as a row 1,2,3,4,5,6,7,8;
1 is right, 2 differs from 1, like from 3, …and so on: equidistant errors!
And now, i need a feature to “train the equidistant errors, such as take into account”; the physical error is x²; a power law. the error delta x will make a power law impact to magnitude…
And this is important to take into account; How can the model learn this, or how to put a hierarchy to this classes? I want to prevent a false -positive/right- because sometimes it is the case that class 4 fits better as class 3, but this is false at all.
i know, this question is like a snake biting itself;
but iam not running out of ideas, there are several approaches:
For example, normalization of the prediction(score):
if the prediction is:
0.953 to class 1
0.922 to class 2
0.897 to class 3
0.865 to class 4
0.833 to class 5
0.809 to class 6
0.780 to class 7
0.751 to class 8
so we can simply see the increasing error…
but it is not equidistant. And sometimes 4 fits better as 3 — false!
And now normalize:
0.953 - 0.922 = 0.031
0.922 - 0.897 = 0.075
0.897 - 0.865 = 0.032
0.865 - 0.833 = 0.032
0.833 - 0.809 = 0.024
0.809 - 0.780 = 0.029
0.780 - 0.751 = 0.029
…
mean = 0.0315
So, now as example class 4… is 0.865. with 0.032 distance to each neighbor.
I can give a deviation of 0.032-0.0315=0.0005, to push it into a kind of equidistance: Normalized prediction range(0.08655 ---- 0.08645) .
This way or another… but the idea is normalization as approach.
Any other ideas? And how to implement?
preventing false-positive through normalization? 
… so there is always a peak in prediction… let us face 15 classes:
0.984 PEAK
0.967
0.945
0.922
0.900
0.968 PEAK(false)
0.867
0.843
0.811
0.793
0.759
0.732
0.708
0.685
…normalization will change the PEAK(false), but this is maybe not suitable.
I cannot manipulate – like going to my own prediction. That makes no sense.
But error management outlier detection should work…
Thank you –
Before select the best fit – i will exclude (prediction)outliers
https://scikit-learn.org/0.20/auto_examples/plot_anomaly_comparison.html
That’s it…
Double MAD - Double Median Absolute Deviation
should prevent anything.