The main purpose of this project is exploration and explanation of boosting algorithms. We want researchers to be able to create and test new algorithms that run in a decent amount of time on their laptops. That code should be understandable by students.
Where we must choose between clarity and performance, we will choose clarity. But we are planning on exploiting Go's capabilities so that our algorithms run on all the cores available :)
If you are a student or researcher interested in Boosting (or Go!), then you might be in the right place. If you're looking for a nice, easy start for writing a test/reference version of your new tweak on boosting, this is probably the place for you.
If you're looking for fast, scalable, distributed, multi-language boosted gradient trees, you almost certainly want XGBoost. It's the choice of professional data scientists everywhere. Our soft goal is be within an order of magnitude of XGBoost's performance on a single workstation. We have no plans to support distributed platforms like Spark and Flink (which XGBoost does)
If you're a Python user or a professional data scientist looking to use boosting as part of a pipeline, you should know that sckikit-learn has implementations of AdaBoost and Gradient Tree Boosting in their Supervised Learning.
If you're looking for reference implementations of decision tree and boosting algorithms in pure Go, then you should probably look at CloudForest instead.
We build with Go 1.7. Earlier version might be fine, but we're not testing against them. The usual go commands (build, test, etc) should work fine.
We use a Makefile and some helper bash scripts (in ./scripts), so you will
need GNU make and Bash to run those. That's no problem for Linux and Mac user,
but if you are on Windows you need to do some extra work. Probably the best
option at this point is the cool new Ubuntu on Windows 10 magic - as described
in lots of places, including
here.
Running make will build and test goboost.
make clean does what you expect.
make test touches the file .tested
make cover will run all tests with coverage and show the HTML report.
Coming - we'll add references as we implement the proper algorithms
Contributions are welcome via pull request. As this project matures, we'll add more guidance about contributing.
See LICENSE for details, but TL;DR it's the MIT license.