LiteRumble

Fragment of a discussion from Talk:RumbleStats
Jump to navigation Jump to search

I just re-implemented the BatchRankings using numpy (which uses C bindings for acceleration), and it also reduced my memory usage from a peak of ~900MB to ~270MB, so no worries any more about ANPP being removed =)

I also added a transform on my APS matrix so that (A)NPP and KNNPBI is perfectly consistent, even if the pairings are not synchronized.

I'm also considering doing a linear regression for the KNNPBI so that we get better interpolation of the expected PBI, particularly at the end points. I'm not sure if this defeats the whole KNNPBI concept though, by introducing 'expectations' based on the APS.

Skilgannon12:57, 28 March 2013