Anyone that built predictors on top of either Kmat...

# datascience

[Hampus Londögård]

Anyone that built predictors on top of either Kmath or Multik? Like SKLearn.

[altavir]

Basic answer is no.

I think that the field is huge, so you need to specify what exactly do you want. Some things are already implemented in Java libraries. For example, regressions are there in Commons-math, and it is quite easy to add a better API for them in KMath (it already has a wrapper for commons-math). Some things need to be implemented from scratch and should be added as a separate module/library.

[altavir]

I am not an ML specialist so I need some kind of input to comment on how to better approach this problem.

[Hampus Londögård]

I agree it's a huge undertaking.

Pinning such a project on Kmath would be incredibly cool if possible, as that would allow us to swap the underlying engine.

But I think it's necessary to allow the Data Science world to move forward on the JVM.

smile is the best library I'm aware of that is "local-first", but the license is a blocker for multiple companies.

[altavir]

If you could open an issue in KMath describing what you want to achieve (not only a reference python, it is huge), it would be a great help. A lot of people are asking what they can contribute.

[altavir]

And since KMath also features Multik integration, it could be done with multik as well.

[Hampus Londögård]

Ah, I now remember the https://github.com/oracle/tribuo project which seems pretty great too.

[Hampus Londögård]

I'll get to work with details. But perhaps it should be a separate project, to clarify the difference between Kmath & that.

A little like numpy & sklearn differentiates.

[altavir]

KMath is an API, and is not monolith. So it makes sense to do such integrations as modules. A common API helps a lot, when you want to integrate different libraries

[Pavel Gorgulov]

I had an idea to make a ML library based on Multik mainly for Android, but it seemed that smile covered the necessary requirements.

[zaleslaw]

Hi, @Hampus Londögård what kind of predictors do you want? What kind of task do you want to solve? Here you can find some examples with solving classical Ml tasks on tabular data with KotlinDL help https://github.com/JetBrains/KotlinDL/tree/master/examples%2Fsrc%2Fmain%2Fkotlin%2Fexamples%2Fml

[zaleslaw]

Also it works with the experimental data frame API https://github.com/zaleslaw/KotlinDL-Demo/blob/master/src/main/kotlin/ml/titanic/Titanic.kt

[Hampus Londögård]

Interesting, thanks!

Depending on TF seems a bit overkill, but indeed interesting. I've implemented my own Logistic Regression / Naïve Bayes for now.

I think building something on top of kmath/multik would be golden to be honest. If we also get DataFrame optimized further Kotlin will be in an amazing place for ML.

[zaleslaw]

Agree with you, @Hampus Londögård.
P.S. TF gives us the ability to train a few types of models (not all, due to C API TF limitations, but a lot of).
Also, at this moment, we could use TF runtime to build classification model (as LogReg, for example) for tabular data on DataFrame