rstan VS LightGBM

I have a Colab notebook with an R kernel that I'm using to share with students for remote lessons in statistics. This notebook relies on "rstanarm", which is pretty massive with the number of dependencies - it takes ~50minutes to install into a fresh Colab session with install.packages(). It seems the issue is that many of the dependencies of this package need to be compiled from source, which takes a long time on Linux distributions like Colab.

You need to install the preview version of rstan: https://github.com/stan-dev/rstan/wiki/Configuring-C---Toolchain-for-Windows

If you're on Windows, there are extra hoop to jump through, I'm afraid https://github.com/stan-dev/rstan/wiki/

And here are the instructions on how to build RStan from source: https://github.com/stan-dev/rstan/wiki

I ran the R code from here

In R, the rstanarm package should do you well. You'll need to install rstan and make sure you have a C++ complier set up as well (instructions here: https://github.com/stan-dev/rstan/wiki/RStan-Getting-Started)

SIRUS.jl is a pure Julia implementation of the SIRUS algorithm by Bénard et al. (2021). The algorithm is a rule-based machine learning model meaning that it is fully interpretable. The algorithm does this by firstly fitting a random forests and then converting this forest to rules. Furthermore, the algorithm is stable and achieves a predictive performance that is comparable to LightGBM, a state-of-the-art gradient boosting model created by Microsoft. Interpretability, stability, and predictive performance are described in more detail below.

Hey, thanks everybody for your answers. I've asked around in the XGBoost and LightGBM repos and some folks there also agreed that memory speed will be a bottleneck yes.

LightBGM seems to have C API support, and C++ example in the main repo

LGBM doesn't do RF well, but it's easy to manually bag single LGBM trees.

LightGBM is a popular and high-performance open-source implementation of the Gradient Boosting Decision Tree (GBDT). To learn how to use this algorithm, please see example notebooks for Classification and Regression.

Microsoft lightGBM. https://github.com/microsoft/LightGBM

Here are our benchmarks on training time comparing Tangram's Gradient Boosted Decision Tree Library to LightGBM, XGBoost, CatBoost, and sklearn.

{ inputs = { nixpkgs = { url = "github:nixos/nixpkgs/nixos-unstable"; }; flake-utils = { url = "github:numtide/flake-utils"; }; }; outputs = { nixpkgs, flake-utils, ... }: flake-utils.lib.eachDefaultSystem (system: let pkgs = import nixpkgs { inherit system; }; lightgbm-cli = (with pkgs; stdenv.mkDerivation { pname = "lightgbm-cli"; version = "3.3.1"; src = fetchgit { url = "https://github.com/microsoft/LightGBM"; rev = "v3.3.1"; sha256 = "pBrsey0RpxxvlwSKrOJEBQp7Hd9Yzr5w5OdUuyFpgF8="; fetchSubmodules = true; }; nativeBuildInputs = [ clang cmake ]; buildPhase = "make -j $NIX_BUILD_CORES"; installPhase = '' mkdir -p $out/bin mv $TMP/LightGBM/lightgbm $out/bin ''; } ); in rec { defaultApp = flake-utils.lib.mkApp { drv = defaultPackage; }; defaultPackage = lightgbm-cli; devShell = pkgs.mkShell { buildInputs = with pkgs; [ lightgbm-cli ]; }; } ); }

I'm working on the AutoML python package (Github repo). In my package, I'm using many different algorithms. One of the algorithms is LightGBM. The algorithm after the training doesn't release the memory, even if del is called and gc.collect() after. I created the issue on LightGBM GitHub -> link. Because of this leak, memory consumption is growing very fast during algorithm training.