Baselines | NVIDIA Structured Data and Graph Models

Baseline Evaluation Metrics

On your training job’s evaluation page, Kumo compares your model against heuristic baselines for temporal binary classification, link prediction, and regression tasks. These baselines show how much your model improves over simple heuristic-based approaches.

Kumo computes baselines at the per-entity level:

Regression tasks use past labels as forecasts.
Link prediction tasks recommend past user behavior as future predictions.

Baseline evaluation metrics panel comparing model performance against heuristic baselines

Kumo does not support baselines when timeframe_step is None.

Baselines Per Task Type

The following table describes the baselines Kumo supports, organized by prediction task type:

Task Type	Baseline
Link Prediction	Per Entity Frequency - Predicts each entity using the most popular ranked list of target values from all previous splits.
Regression	Per Entity Mean (Personalized Average) - Predicts each entity using the average of all past target values from previous splits. Per Entity Last Label - Predicts each entity using the most recent past value as the forecast.
Binary Classification	Per Entity Mean - Predicts each entity using the average of all target values from previous splits. Per Entity Last Label - Predicts each entity using the previous time frame’s label.

Why Use Baselines?

Baselines provide a reference point for measuring how much your Kumo model improves over simple heuristics. A well-performing model should outperform these baselines significantly.