Explanation workflows | Umnai

Explanation workflows show how the platform explanation components can be combined to support higher-level analysis.

The explanation components describe model behaviour at different levels: predictions, attributions, activations, rules, module dependency plots, strengths and weaknesses, histograms, anomalies, and context comparisons. Workflows combine these components to investigate, monitor, justify, optimise, or govern model behaviour.

Workflow areas

Workflow	Description
Model behaviour	Use rules and module dependency plots to inspect trends, behaviour changes, conformance to expectations, and segment behaviour over time.
Identifying bias	Use control swaps, attributions, and module dependency plots to detect bias in predictions or in the model.
Justifying attribution explanations	Use activated rules and their expressions to explain why an attribution was produced.
Identifying data segments	Use rule conditions to understand the regions of data that induction found to behave similarly.
Identifying and addressing low data coverage	Use histograms and activations to find sparse regions, then inspect their behaviour with MDPs and rules.
Model monitoring	Use activation data to detect deviations from expected behaviour, data drift, and potential intrusion or security events.
Model optimisation	Use global explanations such as feature importance and module importance to guide feature selection and module creation.
Real-time explanations	Use the Results view to receive predictions and explanation data at inference time.
Audit logs	Use explanation results as records for transparency, accountability, compliance, and forensic reconstruction.

How workflows use explanation components

Each workflow uses one or more explanation components.

Component	Common use
Attributions	Explain how much features, interactions, or modules contributed to a prediction.
Activations	Identify the partitions and rules used by a query, segment, dataset, or deployed model.
Rules	Show the conditions and local model expressions behind behaviour.
Module dependency plots	Visualise module behaviour and trends across feature values or interactions.
Histograms	Inspect data distribution and coverage.
Strengths and weaknesses	Add confidence context through strength and uncertainty.
Context explanations	Compare a query with swaps, similar cases, dissimilar cases, or related cases.
Results view	Capture prediction and explanation output in real time.

From explanation to action

Explanation workflows can be used to move from inspection to action.

For example, a team may use histograms and activations to find low data coverage, inspect the corresponding module behaviour with MDPs and rules, and then decide whether to collect more data, adjust a workflow, or route affected decisions to a human review process.

Similarly, a team may use control swaps to identify bias in a prediction, compare the attribution deltas, and decide whether to leave the prediction as is, correct the bias automatically, or de-automate the decision for human review.