Try 1000 components of kernel approximation

[nbgl]

Previously used 100 components. Using 1000 will be significantly slower, but may result in improved performance.

[nbgl]

Sure enough, using 1000 components yields a better predictor.

Recall that 100 components gave us

Is the accuracy improvement worth the massively increased computation time?

[nbgl]

This can be closed now. I’ve settled on 1K features.