News for hackers. Stuff that matters.
Re:A Search for Technosignatures Around 11,680 Nearby Stars by ldjkfkdsjnv on Wednesday August 16, 2023 @11:30PM (#37156815)
Reminds me of the failing of the k-nearest neighbor algorithm. It works extremely well in low dimensionality, but given high dimensionality, its astoundingly bad. As you scale up the space and number of dimensions, the distances between individual chunks of data gets so large its hard to draw meaningful conclusions.
Re:A Search for Technosignatures Around 11,680 Nearby Stars by nmca on Wednesday August 16, 2023 @11:40PM (#37156886)
I dunno I've done a lot of high-dim knn that were quite pragmatically useful
Re:A Search for Technosignatures Around 11,680 Nearby Stars by ldjkfkdsjnv on Wednesday August 16, 2023 @11:53PM (#37156976)
Re:A Search for Technosignatures Around 11,680 Nearby Stars by nmca on Thursday August 17, 2023 @11:40AM (#37163378)
Maybe we're talking at cross purposes; what I mean is that PCA+KNN or other tricks mean that for high dim and large amounts of data you can get useful results; with the enlarged view of KNN as "anything where the last step is KNN" your statement that it's "astoundingly bad" is wrong. But for the narrow view of just KNN then sure cod applies.
Re:A Search for Technosignatures Around 11,680 Nearby Stars by natch on Wednesday August 16, 2023 @11:44PM (#37156922)
How high is high?
Re:A Search for Technosignatures Around 11,680 Nearby Stars by ohazi on Thursday August 17, 2023 @12:41AM (#37157314)
I had a professor who would jokingly say "It's lonely in 128 dimensions"
Re:A Search for Technosignatures Around 11,680 Nearby Stars by galaxyLogic on Thursday August 17, 2023 @12:45AM (#37157343)
Right, nobody's at those other dimensions at the same time as you are.