Explanation friendly features

Coming up with features is difficult, time-consuming, requires expert knowledge. “Applied machine learning” is basically feature engineering.
— Andrew Ng, Machine Learning and AI via Brain simulations

How you craft the features that feed your model impacts:

  • learning efficiency
  • model generalizability
  • human interpretation.

Unfortunately sometimes the feature definitions that are most efficient, in terms of run time and resource use, are also features that muddy the waters for human interpretation and projecting generalizability.

Just because you can synthesize a smaller set of features that use less server resource, doesn’t mean you should, especially  if those new features make the black box blacker.  Because the smaller set is statistically equivalent to the larger doesn’t mean they provide the same utility when we consider our broader goals.

When we talk about managing the trade-off between predictive performance and explainability we most often discuss algorithm selection and if additional post-hoc analysis is desired.  Let’s also recognize the role of feature definition in this trade-off.

Good features are:

  • Informative
  • Discriminative
  • Non-redundant

We believe part of being “informative” is preferring features that map to real world attributes, understandable by domain experts, which are inherently more useful in explanations and in judging generalizability.


One thought on “Explanation friendly features

Leave a Reply