# How to improve de-noise algorithm on low signal-to-noise ratio features?

In the following plot, I have features that all have a very small predictive power on y, there is a low signal-to-noise ratio.

In order to de-noise them, I tried PCA and k-mean clustering based on the principal components that are returned by PCA, but the entropy of y on each cluster is still large. Are there better ways to de-noise this kind of data?