Why is my predictive query underperforming for a particular subset of data?

Kumo's column analysis can be incredibly useful for troubleshooting why your predictive query is underperforming for a particular subset of data, as well as discovering bias and areas where you might need to shore up your datasets. By analyzing these charts, you can better understand how individual values within each column positively or negatively affect the final prediction distribution. These statistics are calculated using the ground truth (i.e., the target labels)—which is what the predictive query learns to predict from—as well as the actual predicted values.

If you click on an individual column, you can view a plot that displays that the average model prediction against the actual labels in the holdout data split for different entity populations, segmented based on how often their historic rows appear within certain past time intervals.

• Column Analysis