Model optimisation | Umnai

Model optimisation uses explanation outputs to improve how a model is configured, trained, and reviewed.

Explanations can help identify which features and modules matter most, which ones matter least, and whether those rankings align with real-world knowledge, expectations, and constraints. This can inform both feature selection and the module creation process during induction.

Optimising with explanations

After a model has been induced, global explanations are often used to understand how the model behaves.

Two of the most useful starting points are feature importance and module importance.

These explanations rank and score:

Explanation	What it ranks
Feature importance	Input features, using decomposed attribution values.
Module importance	XNN modules, including feature and feature-interaction modules.

These rankings help identify which parts of the model have the largest and smallest impact on predictions.

Feature selection

Feature importance can help identify low-ranking features.

A low-ranking feature may be a candidate for removal from the training dataset, especially when its low importance aligns with domain knowledge, business expectations, or known constraints.

This should be reviewed carefully. A feature may rank low globally while still being important for a specific segment, interaction, or edge case.

Module creation

Module importance can help identify low-ranking modules.

A low-ranking module may be a candidate for exclusion from the module creation process during induction, especially if the module represents a feature or interaction that is not expected to be useful.

This can help simplify the model and focus induction on structures that are more likely to be meaningful.

Comparing explanations with domain knowledge

Importance rankings should be compared with real-world knowledge, expectations, and constraints.

If a feature or module is highly important but should not matter according to domain knowledge, it may require investigation. If a feature or module is consistently low-ranking and has no expected relevance, it may be safe to remove or exclude.

This comparison helps ensure model optimisation remains grounded in both model evidence and domain understanding.