How to improve the prediction accuracy of a NN / ML model when the correlation between the inputs and output is too low?

Question

Currently, I am working on a regression problem for a production environment where a single output has to be predicted. The data is huge but the problem is that there are almost no useful feature that the model can "learn" to predict the output. My company wants me to build me a model irrespective of the fact, but the accuracy of the prediction is worst which was "expected".

With this limited data as input, I want like to know if there is a way or particular DL / ML architecture which are much suited for a problem like this where the input data to the model have less correlation to the output? What other step can I do from a point of feature generation etc.?

Answer 1

If the features are unrelated to the outcome, then poor performance is the “right” answer.

Think of the regression as a conditional probability distribution. You want to say something about $(Y\vert \vec{X}=\vec{x})$, but if $Y$ and $\vec{X}$ are independent, then $Y=(Y\vert\vec{X}=\vec{x})$.