> 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.

# Debug Model Performance

> Diagnose data, graph, training, and evaluation problems in Kumo AutoML

Debug model performance in a consistent order: verify the task and data first, inspect evaluation behavior second, and tune the model only after the inputs are trustworthy.

## 1. Start with the prediction task

Confirm that the predictive query expresses the intended entity, target, and prediction window. Review the generated label distribution and confirm the task has enough positive, negative, or numeric examples to learn from.

For temporal tasks, verify the anchor time and the ordering of training, validation, and holdout data. Refer to [Anchor Time and Data Leakage](/troubleshooting/what-is-anchor-time-and-why-is-it-important).

## 2. Inspect the data

Begin with a small set of tables and columns, then add information incrementally. Check for:

* Incorrect or inconsistent column types
* Missing or stale values
* Target leakage and future information
* Major distribution shifts between training and holdout data
* Duplicate or mislabeled examples
* Important business signals that are absent from the graph

Refer to [Data Preparation Best Practices](/troubleshooting/data-quality-mistakes-and-how-to-catch-them).

## 3. Check graph connectivity

Confirm that primary-key and foreign-key relationships match the intended schema. Low match rates can indicate incorrect keys or missing related records.

An entity connected only to its own event history receives no signal from similar entities. Add related tables only when they introduce meaningful, well-connected information, and measure whether they improve holdout performance.

## 4. Read the evaluation results

Use the training and validation curves to distinguish common problems:

* **Overfitting:** training performance continues to improve while validation performance levels off or declines.
* **Underfitting:** training and validation performance remain weak, or the curves have not converged.
* **Subset failure:** aggregate metrics look acceptable but performance is poor for a particular cohort.

Use [Evaluation](/evaluation) for metrics and [Explainability](/explainability) to investigate columns, values, and cohorts that influence predictions.

## 5. Compare against a baseline

Compare the model with a simple heuristic, historical baseline, or existing model on the same holdout population and time range. A shared evaluation set prevents data-selection differences from appearing as model improvements.

Refer to [Compare AutoML with an External Model](/troubleshooting/how-can-i-compare-a-predictive-query-with-an-external-model).

## 6. Tune after validating the workflow

After the task, data, graph, and evaluation are correct, consider changing the run mode or model plan. More experiments, a larger search budget, or a longer training schedule can help when the current search has not found a suitable configuration.

Change one factor at a time and compare it on the same validation and holdout data. See [Run Mode](/run-mode) and [Model Plan](/model-planner).