I just had a little interview with Benoit and the subject of missing models came up. After quickly thinking about it, I felt like major classifications models are already implemented. Some classic models are missing such as kNN or Naive Bayes but my gut feeling is that NN and RF are usually more accurate.
However, if I understand well the documentation, the only regression model available (for now) is linear regression. A nice addition would be:
- Lasso and Ridge regressions (or more generally ElasticNet): both are linear models so it is a matter of API (I suppose) since the only difference with Linear Regression is the training which is not handled by Concrete ML.
- Polynomial regression and splines: it might require a bit more work (i.e. not only a simple child class to create) since it is not linear but splines are usually appreciated for their interpretability.
In sklearn, you can find many other regressors but, from my point of view, that’s the two models I would highlight in priority.
By the way, all your work looks awesome! Looking forward to your next releases!