> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.umnai.com/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.umnai.com/_mcp/server.

# Explanation views

Explanation Lab provides local views of a prediction, global views of learned model structure and rules, and tools for generating written interpretations. These guides use Adult Income examples to show how to query a record, inspect its contributions and review the model’s explanations.

## What a view explains

A view presents information derived from a data resource, a trained model or a deployed model. The resource determines what can be inspected: data views show lineage and distributions, model views describe learned structure, and deployed-model views provide predictions, explanations and evaluation results.

In the web app, these outputs appear in the area where you use them. Inspect data in **Data Lab**, evaluate a running deployment in **Model Lab**, and explore predictions and learned behaviour in **Explanation Lab**. See the [Data Lab overview](/ui-guides/data/overview) and [Model Lab overview](/ui-guides/models/overview) for those related workflows.

## Prediction, attribution and model behaviour

A prediction is the output produced for a set of input values. Feature attributions describe how the inputs contributed to that output; module attributions relate it to the model's learned modules, partitions and rules.

A local explanation concerns one query. Global views explore learned structure and behaviour across feature values. A decision rule describes a learned condition; a dependency plot helps show how a module behaves as its inputs vary. These views answer different questions and should be read together when investigating a result.

Evaluation measures performance across a data population. It is separate from explaining an individual prediction: use [Evaluate models](/ui-guides/deployment/evaluation) for classification or regression metrics, and retain the deployment and data selection when comparing results.

## Choose an explanation view

Start with a deployed model and decide what you need to understand. Local views explain a particular set of inputs; global views expose the model’s structure and module behavior.

| **View**              | **Use it to**                                                               |
| --------------------- | --------------------------------------------------------------------------- |
| Local Explanations    | Execute an individual query and inspect feature and module attributions.    |
| Global Explanations   | Browse model structure, module dependencies, and decision rules.            |
| Dependency Graph      | Trace the displayed dependencies from inputs through modules to the output. |
| Module Explorer       | Open the Module Dependency Plot Explorer and adjust its display controls.   |
| XNN LLM Explanations  | Generate a written interpretation of the built-in Adult Income sample data. |
| Symbolic Explanations | Generate a written interpretation of a selected module’s decision rules.    |

## Before you begin

Confirm that the intended deployment is **RUNNING**, check the meaning of its input fields and prepare the record you want to explain. Use a record that is relevant to the task you are investigating.

The examples include zero values and special category labels. Check the input conventions for your model before submitting a query. A displayed default is not necessarily a suitable input for your record.

## Follow a question through the views

1. For one decision, run a local query and inspect its feature and module attributions.
2. For a contributing module, inspect its global decision rules and dependency plot.
3. Use a dependency graph to trace the displayed computation for the same record.
4. If you need written text, review a symbolic explanation against the source rules and retain the context needed to reproduce your interpretation.

The XNN-assisted LLM sample workflow is separate: it interprets supplied sample data and does not inherit the deployed model selected in another view.

## Next steps

Continue to [Local explanations](/ui-guides/explanations/local-explanations).