Start with the explanation question
A local attribution describes how features contributed to a particular model output relative to a reference under a chosen explanation method. A global summary aggregates information across examples and depends on the population being summarized. Neither is the same as proving that changing a feature would cause a real-world outcome to change. First identify the model version, output being explained, units or score space, reference, and intended audience.
A model predicts a delivery time of 40 minutes for one order. State the difference between asking why this prediction differs from a baseline and asking what would happen if the route were changed in the real world.
Interpret baseline and attribution together
In an additive explanation, a baseline output plus contributions approximates or reconstructs the explained output according to the method and output space. For an illustrative regression example, a 30-minute baseline with contributions of +8 and +2 minutes gives 40 minutes. Different baselines or methods can change the contributions. For classification, an explanation may be in a score or logit space rather than probability; inspect the product’s output definition before interpreting values.
A local regression explanation has baseline 30, traffic contribution +8, and distance contribution +2. Calculate the explained output. Explain why the result does not prove that removing traffic would reduce the real trip by exactly eight minutes.
Account for correlations and aggregation
When features are correlated, attribution methods can distribute credit in ways that depend on their assumptions and reference data. A low attribution for one feature does not prove that the underlying factor is irrelevant, and a single explanation does not describe every prediction. Aggregating absolute attributions can summarize typical magnitude but hides direction and variation. Compare relevant slices and inspect individual cases alongside model errors and data provenance.
A model uses both postal area and a correlated regional indicator. If one receives a low attribution, what further checks would you do before concluding that location information has no influence?
Use explanations as investigation evidence
BigQuery ML and Agent Platform provide explanation capabilities for supported models; methods, baseline definitions, and output fields differ. Feature-attribution drift can flag a change in how a model’s inputs contribute to predictions, but it does not alone prove a quality, fairness, or causal problem. Investigate source and schema changes, population mix, prediction outcomes, and relevant slices. Protect explanations from exposing sensitive information and validate access to them.
After an upstream feature update, the attribution distribution changes sharply while model error labels are delayed. List two immediate checks and explain what later evidence is needed to assess model quality.
A feature has a strong positive attribution for one prediction. What is the most defensible conclusion?
Keep exploring
- BigQuery local and global explanations ↗
- Feature attribution methods and limitations ↗
- Monitor attribution drift ↗
Use the official certification page and its linked resources to validate your coverage. For guided labs, explore the Google Skills learning path. Revisit the practical exercise and explain your decision in your own words before moving on.