[WIP] Review: Interpretable Machine Learning

1.sparse high-order interaction model with rejection option (SHIMR)

Maity et al. (2015) paperlink

• Dichotomize variables, and using high-order interaction between dichotomized variables, they find the decision rules.

• Not quite sure why it should be better than partition-based regression like decision tree regression..

• Little down side of the example used: not many subjects, and very easy to classify AD vs. NC, though. They have to compare AD vs. MCI.

2. GoM Model

Das et al. (2019) paperlink

• Not using machine learning, but has some similar flavour with SHIMR kind-of approach.

• This model deals with sex-specific longitudinal memberships, that current machine learning approaches do not deal with. In my personal opinion, AI interpretable should include this kind of interactions, which SHIMR is equipped well to handle if there is any.

3. A good blogs about interpretability

link1

link2

• Importance score (often measured as prediction performance change by permuting a variable)

4. LIME (Local Interpretable Model-agnostic Explanations)

Ribeiro, Singh, and Guestrin (2016) paperlink