MLModelEvaluation
How a trained model scored on rows held out from its training
Fields
Whether the model was scored
Possible values (2)
EVALUATEDThe model was scored on its held-out rows
INSUFFICIENT_ROWSToo few rows to hold any out: the model trained on every row and has no held-out score
How rows were divided between training and the holdout
Rows the model trained on
Rows held out and scored
Rows the split rule could not place (missing key or time value); they trained
Identifies the exact held-out rows; equal fingerprints mean scores on the same rows
The model's scores on the held-out rows
What a trivial model (training mean, majority class, or seasonal naive forecast) scores on the same rows
Classification detail, for a classifier
Regression detail, for a regressor
Backtest detail, for a forecast
Up to 1,000 held-out rows with their predictions, chosen deterministically
The published version and up to four recent others, re-scored on these held-out rows
Referenced by (2)
Exact type SDL
"""How a trained model scored on rows held out from its training"""
type MLModelEvaluation {
"""Whether the model was scored"""
status: MLEvaluationStatus!
"""How rows were divided between training and the holdout"""
split: MLSplitBasis!
"""Rows the model trained on"""
trainRows: Int!
"""Rows held out and scored"""
holdoutRows: Int!
"""
Rows the split rule could not place (missing key or time value); they trained
"""
unsplittableRows: Int!
"""
Identifies the exact held-out rows; equal fingerprints mean scores on the same rows
"""
holdoutFingerprint: String!
"""The model's scores on the held-out rows"""
metrics: TrainingMetrics
"""
What a trivial model (training mean, majority class, or seasonal naive forecast) scores on the same rows
"""
baseline: TrainingMetrics
"""Classification detail, for a classifier"""
classification: MLClassificationEvaluation
"""Regression detail, for a regressor"""
regression: MLRegressionEvaluation
"""Backtest detail, for a forecast"""
forecast: MLForecastEvaluation
"""
Up to 1,000 held-out rows with their predictions, chosen deterministically
"""
samples: [MLHoldoutSample!]!
"""
The published version and up to four recent others, re-scored on these held-out rows
"""
comparisons: [MLVersionComparison!]!
}