Justifying attribution explanations | Umnai

Attribution explanations have two parts: how and why.

Attributions explain how a prediction was built. They show how much each feature, interaction, or module contributed to the prediction. Activated rules explain why those attributions were produced by showing the conditions and local behaviour behind them.

How and why

When explaining a prediction with attributions, the attribution values explain:

Part	Explanation
How the prediction was built	The prediction is built from the sum of its attributions.
How much each component contributed	Each feature, interaction, or module has an attribution value.

The activated rules explain why those attribution values were produced.

Each activated rule contains:

Rule expression	Role
`conditional_expression`	Identifies the data segment or group that triggered the rule.
`then_expression`	Describes the behaviour of that segment and how the attribution is calculated.

In many cases, the conditional_expression is enough to justify the attribution. When more detail is needed, the then_expression can be used to explain the behaviour of the segment.

Example

Suppose the Age module attributed 0.15 for a query where Age = 10.

The activated rule in the Age module is:

IF (0 < Age ≤ 18) THEN attribution(Age) = 0.05 + (0.01 * Age)

The attribution can then be justified at different levels of detail.

Basic justification

Age attributed 0.15 because Age is between 0 and 18.

This uses the conditional_expression to explain which segment triggered the attribution.

Detailed justification

Age attributed 0.15 because Age is between 0 and 18, and the behaviour of that segment follows the equation 0.05 + (0.01 * Age).

This adds the then_expression, which explains how the attribution is calculated for the activated segment.

Human-friendly justification

Age attributed 0.15 because Age is between 0 and 18, and for that age group the attribution increases linearly with Age, from 0.05 up to 0.23.

This keeps the same information but presents it in a more natural way for a non-technical reader.

When to use each level

Different audiences need different levels of justification.

Level	Best for
Basic justification	Explaining the segment or condition that caused an attribution.
Detailed justification	Debugging, audit, technical review, or model validation.
Human-friendly justification	User-facing explanations, operational workflows, or business review.

The important point is that each level is grounded in the same activated rule. The explanation can be shortened or translated into friendlier language without losing its connection to the model structure.