[mlpack-git] [mlpack/mlpack] Vantage point tree (#708)

[Ryan Curtin]

Oops, it seems like I did not look closely enough at the traversers to see how they were different, so it seems like most of my comment was not necessary... sorry about that.

I thought of something else, that may allow us to avoid needing a `Score(queryNode, referenceIndex)` function in each rules class.  What if each node in the vantage-point tree had three children: a zero-radius child that only contained the vantage point, the inner child (that does not contain the vantage point), and the outer child?  The `MinDistance()` and `MaxDistance()` functions would be correct for the vantage point child, so there would be no need for a `Score(queryNode, referenceIndex)`.  I think also that might simplify traversal greatly, because the vantage point child (the zero-radius child) will be a leaf node, and thus will be kept during the traversal.  Clever ordering of which child to visit first would result in, I think, an identical traversal to what you've already implemented.  Let me know what you think of that idea---like most of my ideas, I have not thought through the full details yet in order to try and provide a quick response, so it's possible there are fatal flaws I have not encountered in my thinking.

I haven't thought completely about why the pruning might not be doing a good job.  I'll try and think about this in the next few days, but I suspect that the `NeighborSearchRules` as implemented aren't actually doing the same thing (or even a similar thing) to the algorithm given in the paper.  I'll let you know what I think of.

---
You are receiving this because you are subscribed to this thread.
Reply to this email directly or view it on GitHub:
https://github.com/mlpack/mlpack/pull/708#issuecomment-234580961
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack-git/attachments/20160722/0ec36041/attachment-0001.html>