A Julia machine learning environment combining convenience and control, through a combination of high and low-level interfaces. Currently in development and experimental.
A common high-level interface is provided through the systematic
implementation of default pretransformations of training and testing
data. In particular, all learning algorithms receive learning data in
a common format (a DataFrame object) with all algorithm-specific
transformations occurring under the hood. This allows for the quick and
efficient comparison of several learning models. To mitigate against
data leakage, data transformations are "fit" on training data
only. However, there is a provision for dealing automatically in
testing or (cross) validation with classes of categorical variables
not seen during the fit, a common annoyance.
For an introductory tour, clone the repositories into directories that your Julia installation can find and run docs/tour.jl in your Julia REPL. Note that Koala and associated libraries require Julia 0.6.
KoalaTrees: Supervised learning with single regularized trees
KoalaEnsembles: Build weighted ensemble learners (e.g., random forests, extreme random forests)
KoalaLightGBM: A Koala wrap of Microsoft's gradient tree boosting algorithm
KoalaElasticNet: The elastic net and lasso linear predictors
KoalaRidge: Ridge regression and classification
KoalaKNN: K-nearest neighbor
KoalaTransforms: A library of common data transformations (and dependency of several of the other libraries)
KoalaFlux (coming soon): A wrap of Mike Innes' beautiful Julia implementation of neural networks, including a facility to learn categorical feature embeddings.
At present the above implement supervised regression (todo: classification). To learn how to wrap your favorite machine learning code for use in Koala, refer to:
KoalaLow: To expose Koala's low-level interface