K nearest neighbor utility in Julia

Community
1

Hi. Is there a working K-nearest neighbors package or sub-package available in Julia (apart from Scikit Learn)? There used to be kNN but that is not currently working. A search within the past year yields essentially nothing (except a proposed interface). Maybe it exists as a part of another package? Thanks for any help.

2

GitHub - KristofferC/NearestNeighbors.jl: High performance nearest neighbor data structures and algorithms for Julia. might suit your needs.

5 Likes
3

Thanks very much for that. I am looking for something for elementary students in applied Data Science to use, with a simple interface. This package certainly looks very powerful, and potentially useful for academics and serious researchers! I was hoping for something like the (defunct) kNN.jl package.

Perhaps I can write them a wrapper using this, that doesn’t require them to understand trees, etc.

4

Yeah it doesn’t look like rocket science ( imho it’s never bad to understand the parts that make a method work), but you could easily set up an API similar to scikit where fit builds a tree and predict does the search over input points.

5

Thanks, yes, in this class we are reviewing and comparing the performance of lots of methods, and don’t have time to go into depth, unfortunately. I probably can write something, but I wondered if something already existed (as I mentioned, the old kNN.jl would have been perfect).

I agree it’s not Rocket science, and am glad of that, because it means that even I might be able to understand it well enough to write the wrapper!

6

They don’t have to “understand” trees. They just call one constructor and then the knn function. I think they can handle it :slight_smile:

8 Likes
7

I still think there are quite a few non-obvious steps involved if one wanted to use this for kNN classification/regression (it would take me a day to figure it out and debug it).

8

MLJ has kNN classifiers and regressers based on NeatestNeighbors.jl

In general it’s a go-to for “classical machine learning” (and more)

1 Like
9

Okay, it seems that you are not really interested in just doing a knn search (which was my initial assumption) but you want a more full fledged classifier framework. Then yes, you probably have to look “higher up” the stack.

10

MLJ has kNN classifiers and regressers based on NeatestNeighbors.jl

In general it’s a go-to for “classical machine learning” (and more)

Thanks for that. I thought MLJ was only Neural Networks, etc. (I quite like Flux), but I’ll definitely take a look! This might be what I was hoping to find.

11

Hi. Thank you for your input (and also for the package, which seems to drive the suggestion I will probably go with) I’m sorry I didn’t make my ‘Data Science’ requirement clearer above.

12

Some nearest neighbor implementations for various data can be found here: https://github.com/JuliaNeighbors

13

MLJ is actually mostly a tool to use and compose “general ML” models though it does in particular interface Flux via MLJFlux. Otherwise it exposes most models from ScikitLearn as well as a bunch of Julia models like NearestNeighbors, DecisionTrees, GLM, LightGBM, MLJLinearModels, …

4 Likes
14

Edit: I read too fast you message, I didn’t implemented the KNN algorithm (not yet…), altought it is discussed in the notes cited below. (here)… If you want to collaborate in writing it… :wink:

I believe you could find interesting the cluster module of BetaML… I have wrote it with exactly your user case in mind: something simple for learning the algorithm and a easy to read code, altought not as much performant as in other packages …

Note that the algorithms in BetaML have a companion repository of the notes of a MITx course in machine learning where the algorithms are explained in greater detail…

1 Like