Global explanations | Umnai

Global explanations describe an Umnai model as a whole.

They explain the structure of the XNN, the patterns discovered during induction, and the components that influence predictions across a dataset or segment.

Global explanations answer the question:

How does the model make predictions?

They are different from local explanations, which focus on a specific query or batch of queries. Global explanations help you understand the model’s overall behaviour before inspecting individual predictions.

Why global explanations matter

XNNs are interpretable because their internal structure is inspectable.

Modules, partitions, rules, bins, attributions, dependencies, constraints, and dataset characteristics can all be used to understand how a model behaves. Global explanations organise that information so the model can be reviewed at the dataset, segment, feature, module, and interaction level.

This is useful for model review, monitoring, validation, bias analysis, and understanding whether the model has learned behaviour that matches the data and domain context.

Global explanation components

Global explanations include several components.

Component	Description
Dataset	Dataset-related information used to understand the data behind the model, including features, histograms, anomalies, segments, dependencies, and constraints.
Module and feature importance	Global attribution summaries that show which modules, features, and interactions have the largest average impact on predictions.
Module dependency plot	A visualisation of how a module’s attribution changes with respect to its input features.
Strengths and weaknesses	Confidence-style measures that describe model strength and uncertainty across activated partitions, data segments, batches, or datasets.

Dataset explanations

Dataset-related explanations describe the data used to create and understand the model.

They include feature information, engineered features, histograms, feature groups, anomalies, windowing, segments, time-series transformations, dependencies, and constraints.

These concepts help explain not just what the model learned, but also the data context in which it learned it.

Importance explanations

Module and feature importance explain which parts of the model matter most at a global level.

Module importance ranks XNN modules by their average absolute attribution across a dataset or segment. This shows which features and feature interactions have the largest impact on predictions.

Feature importance uses decomposed attributions to rank individual features by their average absolute contribution across multiple rows.

Dependency explanations

Module dependency plots show how a module’s attribution changes as its inputs change.

They are similar in purpose to partial dependency plots, but are generated directly from the Hybrid Intelligence model. This makes them a direct representation of what the model learned about a feature or feature interaction.

Strength and uncertainty explanations

Strengths and weaknesses describe confidence-related behaviour at a global level.

Strength reflects the amount of training data in activated partitions and the quality of the partition model fit. Uncertainty reflects estimator variance and partition prediction variance. These measures can be aggregated across a segment, batch, or dataset.