You are an ML engineer at a bank. You have developed a binary classification model using AutoML Tables to predict whether a customer will make loan…

The request is to explain the model's decision for a single customer. This requires a local explanation, which details the factors contributing to an individual prediction. AutoML Tables provides local feature importance (also called feature attributions) for this exact purpose. This feature shows how much each feature value for that specific customer contributed to the final prediction, directly answering why the loan was rejected. Global methods, in contrast, describe the model's overall behavior, not the reasoning for a specific case.

B. Use the correlation with target values in the data summary page. This is a global statistic about the entire dataset, not an explanation for a single prediction.
C. Use the feature importance percentages in the model evaluation page. This is a global feature importance metric, indicating which features are most influential for the model overall, not for one specific customer's prediction.
D. Vary features independently to identify the threshold per feature that changes the classification. This is a manual and less reliable method of sensitivity analysis. It fails to account for feature interactions and is less systematic than the built-in local importance feature. ---

1. Google Cloud Documentation, "Get online explanations": "For AutoML Tables models, you can get both global feature importance and local feature importance (feature attributions). Global feature importance tells you how much each feature impacts a model overall. Local feature importance tells you how much each feature impacted a particular prediction." This directly supports using local importance for a single prediction.

Source: Google Cloud, Vertex AI Documentation, "Get online explanations".

2. Google Cloud Documentation, "Feature attributions": "Vertex AI Explainable AI helps you understand your model's outputs for classification and regression tasks. [...] For a given prediction, feature attributions indicate how much each feature contributed to the prediction. [...] When you request an explanation, you get the predicted class along with the feature attributions."

Source: Google Cloud, Vertex AI Documentation, "Overview of Vertex AI Explainable AI".

3. Lundberg, S. M., & Lee, S. I. (2017). A Unified Approach to Interpreting Model Predictions. This foundational paper introduces SHAP (SHapley Additive exPlanations), a game-theoretic approach that is a primary method used by Google's Explainable AI to calculate local feature importance. The paper's focus is on explaining individual predictions.

Source: Advances in Neural Information Processing Systems, 30 (NIPS 2017), Section 3, "Additive Feature Attribution Methods". DOI: Not available for NIPS proceedings, but accessible via https://proceedings.neurips.cc/paper/2017/hash/8a20a8621978632d76c43dfd28b67767-Abstract.html