Explanation conceptsLocal explanations

Decision action table

Understand rules, activations, and attributions in a single local explanation table.
View as Markdown

The Decision Action Table, or DAT, brings the key components of a local explanation into a single table.

It is designed to make the main components of each decision easier to inspect: the activated rules, the conditions that triggered them, the resulting actions, and the attribution values associated with those actions.

What the DAT contains

The DAT provides a detailed view of rules, attributions, decompositions, and activations for a decision-making model based on an Explanation Structure Model (ESM).

It can include:

ComponentDescription
Module attributionsContributions from each feature or feature interaction module.
Feature attributionsContributions grouped by individual input feature.
DecompositionsFeature-level components of interaction attributions.
Activation dataActivated partitions and rules for the query.
Activated rulesThe conditional_expression and then_expression for activated rules.

Decision and action

The DAT combines two perspectives: decision and action.

The decision part of the table breaks down the conditional_expression for each bin of each rule in each partition of the model. It shows which partitions were activated for a query and why those partitions applied.

The action part of the table shows the decomposed attribution or activation from each activated bin. This connects the condition that activated with the contribution that resulted from it.

Together, these views make it possible to trace a prediction from query conditions to activated model structures and attribution values.

Why the DAT is useful

The DAT is designed to support detailed local explanation work.

UseDescription
Analyse decisionsBreak down conditions and contributions for a query.
Explain activationsShow which partitions and rules activated, and why.
Justify actionsConnect activated conditions to their associated attribution values.
Support expanded and non-expanded conditionsChoose the representation that best fits the analysis or visualisation.
Filter local explanationsFocus on the modules, features, or conditions most relevant to a query.

Filtering the DAT can provide a quick summary of the conditions associated with a query and the resulting actions.

Examples

DAT for query with CapitalGain = 0 filtered to show CapitalGain module only

DAT for query with CapitalGain = 0 filtered to show CapitalGain module only.

DAT for query with Age = 35 and Education = Assoc-voc filtered to show all Age-related modules

DAT for query with Age = 35 and Education = Assoc-voc filtered to show all Age-related modules.