GCP Professional ML Engineer Practice Question 10: Explanation Method Differentiable Model

Integrated gradients is the recommended method for differentiable models such as neural networks, and Google's guidance calls it out specifically for models with large feature spaces; it supports classification and regression on tabular data. Because it evaluates gradients at intervals along an integral path rather than re-running the model over many feature subsets, its cost scales with the number of path steps rather than with the number of features, which is what makes 400 features at production volume affordable. It also carries the same axiomatic properties as the Shapley value, so the attributions are defensible to a regulator.

Option A is the right family of answer for the wrong model type. Sampled Shapley is recommended for non-differentiable models, ensembles of trees and neural networks, which is why AutoML tabular models use it, and its sampling over feature permutations is exactly the cost that blows up with a 400-feature input at this request volume.

Option B is for images. XRAI builds on integrated gradients but adds oversegmentation and region-level saliency maps; there are no image regions in a tabular credit application.

Option D is a monitoring objective, not an explanation method. Feature attribution monitoring watches how attributions shift over time in aggregate; it produces no per-decision reason codes for an individual declined application.

Reference: Introduction to Vertex Explainable AI