Strengths and weaknesses | Umnai

Strengths and weaknesses, or SAW, provide confidence-related measures for a local explanation.

For each prediction, SAW returns two measures: strength and uncertainty. Together, they help describe how well supported the prediction is by the activated parts of the model, and how stable or uncertain those activated parts appear to be.

Strength

Strength reflects model strength for the activated partitions behind a prediction.

It is based on:

Factor	Description
Training data support	The amount of training data in each activated partition.
Partition model fit	The quality of the partition’s polynomial model, or `then_expression`, fit.

A prediction is stronger when its activated partitions have better data support and better local model fit.

Uncertainty

Uncertainty reflects model certainty for the activated partitions behind a prediction.

It is based on:

Factor	Description
Estimator variance	Variation across estimators.
Partition prediction variance	Variation in the predictions produced by the partition’s polynomial model.

Higher uncertainty indicates that the model behaviour is less stable or less certain for the activated partitions that contributed to the prediction.

How SAW is calculated for a prediction

A query activates one partition in each module.

The strength and uncertainty values for a prediction are calculated by aggregating the measures across those activated partitions. This produces prediction-level confidence signals that can be interpreted alongside the prediction, module attributions, feature attributions, and activation path.

How to use SAW in a local explanation

SAW helps add confidence context to a prediction.

Signal	Interpretation
High strength	The prediction is supported by stronger activated partitions.
Low strength	The prediction may rely on weaker activated partitions.
Low uncertainty	The activated partitions appear more stable or certain.
High uncertainty	The activated partitions appear less stable or more uncertain.

SAW does not replace attribution. Attributions explain what contributed to the prediction. SAW helps describe how strong or uncertain the supporting model behaviour is.