Task Title
Task: Automate selection of number of features
Task Description
Part of #134 but can be delivered independently and would already bring great value in allowing to re-fit models automatically.
cfr message to Nick of 21/3:
Hey Nick,
Benoît is right, you can just script one run of your model creation for future re-fits.
The only big hurdle indeed is the selection of parameters. If you fix the number of features to select to a reasonable number (for example, keep it equal to the currently selected number of features), the client can recurrently re-fit the model.
This solution is a bit sub-optimal though, in that new data or even new features may result in the model under- or over-fitting.
The manual process we make as data scientists to select the optimal number of parameters though, is easy, just an elbow curve problem. Algorithms already exist to select the optimal number automatically - mainly on elbow curves for clustering (or you could implement a simple check yourself on how the slope against the elbow curve evolves at each newly added feature). This will could be solved on the project of the client you mention, +- 2 days of work.
Task Title
Task: Automate selection of number of features
Task Description
Part of #134 but can be delivered independently and would already bring great value in allowing to re-fit models automatically.
cfr message to Nick of 21/3:
Hey Nick,
Benoît is right, you can just script one run of your model creation for future re-fits.
The only big hurdle indeed is the selection of parameters. If you fix the number of features to select to a reasonable number (for example, keep it equal to the currently selected number of features), the client can recurrently re-fit the model.
This solution is a bit sub-optimal though, in that new data or even new features may result in the model under- or over-fitting.
The manual process we make as data scientists to select the optimal number of parameters though, is easy, just an elbow curve problem. Algorithms already exist to select the optimal number automatically - mainly on elbow curves for clustering (or you could implement a simple check yourself on how the slope against the elbow curve evolves at each newly added feature). This will could be solved on the project of the client you mention, +- 2 days of work.