Forward and backward testing of contextual effects

[cnellington]

Typically context dependence is understood through univariate contextual effects, which describes the individual contribution of a single feature on a model's parameters. Alternatively, contextualized modeling jointly accounts for many contextual effects without requiring users to specify a set of known contextual dependencies, but these features' individual effects may also be context dependent through synergistic or interaction effects.

Drawing a connection between testing for multivariate and univariate contextual effects would relate effects discovered by contextualized.ml to well-established tests for univariate contexts.

We might consider constructing contexts feature-wise using sequential rounds of univariate testing (forward). We might also attempt to test for the most meaningful contextual features from a contextualized model that has been trained on all contexts at once (backward).