Expensive information? Please. I'm not asking for all weights, I'm asking for 1/1000 weights. If it's tiny, then sharing it should be inconsequential. And no bullshit about competitors or spammers please -- it's not as if anyone is unaware that ranking well in Google organic results is good :rolleyes:
I thought it went without saying that the relevant coefficient should be taken from a unit normal vector or some other indication of relative magnitude should be given.
I prefer expectation (over typical queries) of weighted variance of the feature, divided by the total weighted variance of all features. (most likely sqrt(variance) instead, but whatever power you prefer).
