Skip to main content
Version: Next

Segment performance

Overview​

A segment performance test lets you explore a model's data subsets (segments) that diverge from average scores. In other words, a segment performance test allows you to discover which data points (segments) the model struggles, outperformance, and performs with when generating accurate predictions.

To run a segment performance test on a model, H2O Model Validation utilizes a provided dataset to generate model predictions to assess their accuracy. H2O Model Validation splits the dataset into segments by the bins of values of every variable and every pair of variables to generate results around the ability of the model to produce accurate predictions with different data segments. These results are embedded into a bubble graph that H2O Model Validation generates that enables you to observe and explore data segments the model struggles or performs well when generating accurate predictions. For each segment, H2O Model Validation calculates its size of it relative to the size of the dataset and estimates the error the model makes on the corresponding segment.

H2O Model Validation provides several settings for the segment performance test.

Bubble graph

Exploring and identifying data segments in a model that do not perform as expected can lead to decisions on how you preprocess your data.

Resources​

Feedback