Is there an efficient approach in machine learning when I have the confidence (uncertainty) values for the input features?

Question

Could you give me some comments?

I'm looking for a better approach when I have confidence (uncertainty) values for each input feature.

For example, let's say each class has 3 features.

f1,value = 0.003
f2,value = 0.005
f3,value = 0.007

And they have their own confidence value [0-1 range]. (It means how reliable the feature is.)

f1,conf = 0.2 (not reliable)
f2,conf = 0.8 (reliable feature!)
f3,conf = 0.6 (so-so)

Actually, the low confidence value is because of noise and other interferences.

I'm using conventional neural network and put these values in series.

ex) [f1value, f1conf, f2value, f2conf, f3value, f3conf]

But is there any better approach to use the confidence values?

For example, is there any way to put 'confidence values of input features' to learning network?

----added

Actually, I have the confidence values. You can think of these values as signal quality.

Answer 1

First of all I think that including confidence values as features doesn't make sense from a ML perspective. The algorithms themselves are designed in a way to perform feature selection. However, if you want to compute confidence values for each feature (meaning how informative the feature is) you can use a statistical test to compute that. An example for such test is Pearson's chi-squared test. When it comes to the implementation you can have a look at the Sklearn's SelectKBest. If you pass the chi-squared score function (or any other feature selection metric) you can return the feature importance scores.