Authors : Rémy Deshayes and Sarah Lauzeral
This is an early-stage shot at implementing a Parallel Lasso with Cython.
Over the years, the amount of available data has sky-rocketed to the extent that some training datasets cannot be stored in a single machine memory.
In that case, traditional subgradient methods such as the Coordinate Descent algorithm - focus of this project - can no longer be used. Indeed, Python and R packages often kick off by loading the entire dataset on the RAM.
[1] Distributed Coordinate Descent for L1-regularized Logistic Regression, Ilya Trofimov, Alexander Genkin (2015)