Dimentionality Reduction before machine learning about cosmic-rai HOT 3 OPEN

I'm looking at the various tools in Sklearn's feature_selection module - they should be able to help with this problem as well.

from cosmic-rai.

I tried PCA to reduce the 100+ sensors down to n=3 or n=10 attributes, which are supposed to reflect the n-axis that explain the most variance. But they don't seem to explain much variance. I think we have to be a little careful about choosing techniques that work well with sparse data, if we wanna reduce charge features. What are you thinking about?

from cosmic-rai.

I tried basically sorting the sensors for each event and picking the top 10 or 20. I was thinking this could be a naive form of cropping-- we lose the true structure of the sensors, but maybe there is still a distinctive pattern. Bad results.

from cosmic-rai.