Understand Explanations

Use Kumo Relational explanations to investigate which historical examples and related records are most strongly associated with a prediction. Kumo Relational provides two complementary views: cohorts group historical examples with known outcomes, called context examples, while local subgraphs show related data for one prediction entity, the entity being scored. These views can help you identify patterns and data-quality issues, but they do not establish why an outcome occurred or what caused it. Refer to Prediction Explainability to configure explainability and request an explanation.

Cohort analysis

Cohorts group sampled context examples by shared values and summarize how each group is associated with model targets or scores. A cohort can include the source table and column, the number of relationships traversed from the prediction entity (graph-hop distance), metadata that tells Kumo Relational how to interpret the column (semantic type), value buckets, population share, and average target.

Large differences between cohort targets can reveal a pattern worth investigating. However, they do not establish that changing a column will change the outcome.

When comparing cohorts:

• Confirm that each bucket has enough examples.
• Compare the same column at the same graph-hop distance.
• Check whether missing values form a distinct pattern.
• Repeat the analysis across related entity samples and anchor times, the points in time from which predictions are made.

Local subgraph attribution

A local subgraph is the subset of connected rows, cells, and relationships sampled around the requested entity. Attribution values indicate which parts of this sampled data are most strongly associated with the model output.

Node IDs are internal indices, not source primary keys. When presenting an explanation, preserve the source table, key values, and anchor time so that a reviewer can trace the explanation back to the source data.

Avoid universal labels such as “negligible,” “strong,” or “critical” for numerical attribution ranges unless the deployed explanation method defines and validates the corresponding thresholds. Compare scores within the same explanation, and use domain review to determine whether an attribution is material.

Context examples

When present, the context_examples list in a subgraph identifies the historical examples associated with the explanation. Each entry can include an entity ID, a seed time that identifies the example’s reference time, a score, and a known outcome label.

Compare scores only within the returned explanation. A higher score does not establish that an example caused the prediction or that it will remain the closest comparison across different explanations. Check whether high-scoring examples are valid, sufficiently recent, and similar in ways that are relevant to the task. Unexpected examples can reveal problems with the graph, timestamps, or semantic types.

Natural-language summaries

Generated summaries make structured explanation data easier to scan, but they can omit uncertainty or overstate associations. Keep the underlying cohorts, subgraphs, and scores available to reviewers. Verify every factual statement in a summary against the structured explanation data.

Interpretation checklist

• Are the prediction entity and anchor time correct?
• Were influential rows available by the anchor time?
• Are the cohort sizes large enough to interpret?
• Does the pattern persist across reruns or related entities?
• Could a sensitive or proxy attribute be driving the association?
• Can a reviewer trace the claim back to structured explanation data?

Use explanations responsibly

• Treat cohorts, context-example scores, and subgraph attributions as model associations, not causal effects.
• Compare explanations across reruns, anchor times, and related entities before relying on a pattern.
• Protect sensitive values that may appear in subgraphs or generated summaries.
• Do not use an explanation as the sole justification for a high-stakes decision.

Interpret explanations alongside the underlying graph, query, entity, and anchor time.