ANN search is known to outperform NN search in terms of efficiency and some techniques reduce storage space from compact representations. But what happens in terms of effectiveness? Is it possible to achieve the same performance without finding the nearest neighbor with exhaustive search?
Asked
Active
Viewed 71 times
2 Answers
1
If by effectiveness, you mean accuracy (i.e. finding the exact nearest neighbor), then no. NN search will always find the exact NN, while ANN search will, at its best chance, find the exact NN, that is a tie at the result with NN search.
However, in a high dimensional space the curse of dimensionality lurks and the usual data structures and algorithms for 2D and 3D tend to be as slow as the brute force search, thus an ANN search is the way to go, when you (big) data live in a high dimensional space.
gsamaras
- 71,951
- 46
- 188
- 305
-
I agree with you but not 100%. When we speak of exact NN, this does not have to be the real NN, our accuracy is still not 100%. This is because the search is influenced by the representation used for our data. It may be possible to change the focus of the question to determine if it is possible, through ANN search, to eliminate false positives that affect our accuracy. – jperezmartin Dec 23 '16 at 16:47
-
Well, if you have compressed your data, then of course the representation may affect the accuracy, but that's not a problem of the exact NN. I mean that's actual ANN, but it all depends on the view point! @jperezmartin Hope the answers helps! =) The posted question was nice BTW! – gsamaras Dec 23 '16 at 19:53
0
I tried binary search and ann search on ip2location database. It has the same speed but with many optimizations. You can find a source code at https://ip2locationphp.codeplex.com/.
Micromega
- 12,486
- 7
- 35
- 72