Try 1000 components of kernel approximation

chengsoonong / mclass-sky

Multiclass methods for astronomical data

BSD 3-Clause "New" or "Revised" License

9 stars 4 forks source link

Open nbgl opened 7 years ago

nbgl commented 7 years ago

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

nbgl commented 7 years ago

Sure enough, using 1000 components yields a better predictor. figure_1

Recall that 100 components gave us 100components

Is the accuracy improvement worth the massively increased computation time?

nbgl commented 7 years ago

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