Explanations

Explanations expose how the platform models make predictions.

They connect model outputs to the structures behind them: features, modules, partitions, rules, attributions, activations, strengths and weaknesses, and contextual comparisons. Together, these explanation types make model behaviour inspectable at different levels of detail.

Explanation types

The platform explanations can be grouped into three broad types.

Global explanations

Global explanations describe the XNN model as a whole and the patterns that induction identified in the data.

They help explain model behaviour across a dataset, rather than for one prediction only. Global explanation components include dataset-related explanations, module and feature importance, module dependency plots, and strengths and weaknesses.

Local explanations

Local explanations describe predictions for a query or a batch of queries.

They help explain why a model produced a specific prediction while keeping that prediction connected to the global structure of the XNN. Local explanation components include predictions, module attributions, feature attributions, activations, decision action tables, and strengths and weaknesses.

Local explanations can also be compared between queries.

Context explanations

Context explanations enrich local explanations by adding comparison or scenario context.

They can be used for checks such as what-if analysis, control swaps, and nearest-neighbour comparisons. These explanations help show how a prediction or attribution would change under a different condition, or how it relates to similar, dissimilar, or related cases.