Bringing a Ruler Into the Black Box: Uncovering Feature Impact from Individual Conditional Expectation Plots
This work addresses the need for interpretable feature impact metrics for practitioners using complex models, representing an incremental improvement over existing visual methods like PDPs and ICE plots.
The paper tackles the problem of interpreting feature impact in machine learning models by proposing ICE feature impact, a metric derived from Individual Conditional Expectation plots that quantifies feature influence analogously to linear regression coefficients, and demonstrates its utility on real-world data.
As machine learning systems become more ubiquitous, methods for understanding and interpreting these models become increasingly important. In particular, practitioners are often interested both in what features the model relies on and how the model relies on them--the feature's impact on model predictions. Prior work on feature impact including partial dependence plots (PDPs) and Individual Conditional Expectation (ICE) plots has focused on a visual interpretation of feature impact. We propose a natural extension to ICE plots with ICE feature impact, a model-agnostic, performance-agnostic feature impact metric drawn out from ICE plots that can be interpreted as a close analogy to linear regression coefficients. Additionally, we introduce an in-distribution variant of ICE feature impact to vary the influence of out-of-distribution points as well as heterogeneity and non-linearity measures to characterize feature impact. Lastly, we demonstrate ICE feature impact's utility in several tasks using real-world data.