MLJ is a popular machine learning framework for Julia. It ships with a large suite of common machine learning models, both supervised and unsupervised. It seems natural to interface this package to MLJ, although currently differentiability is a major challenge: to be able to use any of our counterfactual generators to explain MLJ models, those models need to be differentiable with respect to features. Still, this is worth exploring.
I propose the following steps:
[ ] Implement basic interface to MLJ (essentially have an AbstractFittedModel for MLJ.Supervised)
[ ] From the MLJ model list, identify which ML models fulfil the differentiability criterium. Note that some models, like decision trees, may be differentiable after probability calibration
pat-alt
commented
1 year ago
MLJ is a popular machine learning framework for Julia. It ships with a large suite of common machine learning models, both supervised and unsupervised. It seems natural to interface this package to MLJ, although currently differentiability is a major challenge: to be able to use any of our counterfactual generators to explain MLJ models, those models need to be differentiable with respect to features. Still, this is worth exploring.
I propose the following steps:
AbstractFittedModelforMLJ.Supervised)