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

# How does explainability work?

> Understand column sensitivity, feature cohorts, and sampled graph context

Kumo Engine generates explanation files after fitting a model, when the task and model support them:

* **Gradient-based column scores** summarize sensitivity to input features by table and column.
* **Column cohorts** group feature values into ranges or categories and report population shares, average predictions, and observed targets when available.
* **Sampled graph context** records nodes, relationships, and feature values for selected examples.

These describe the model on the test examples analyzed. They do not establish causality or explain every later prediction.

`Trainer.fit()` requests explanations unless `trainer.model_plan.training_job.disable_explain` is enabled. Files are written under the selected experiment, or its training-and-validation refit when present. Save them outside scratch if you need to retain them; they are not included in a published model bundle.

Classification and regression are eligible for all three methods. Link prediction supports gradient scores and sampled graph context; forecasting supports gradient scores. Multilabel ranking is not supported by these methods.

## Missing Explanation Scores

A missing score does not mean the column was excluded from training. Check task and model support, `disable_explain`, and fit logs. An explanation step can fail without failing the fit.

Inspect the selected model's encoder settings to confirm feature inclusion. A `Null()` encoder excludes a column from model features but preserves its key relationships in the graph. See [Encoder Selection](/kumo-engine/user-code-guide/set-up-a-graph/encoder-selection).

See [Model Explainability](/kumo-engine/user-code-guide/train-a-model/model-explainability) for the support matrix and code to locate and read the files.