- Respects the input and output specifications and the performance requirement
- Benefits the organization (measured via cost-reduction, increased sales or profit)
- Helps the user (measured via productivity, engagement, and sentiment)
- Is scientifically rigorous (predictable and reproducible)
#Predictability
If the input feature vectors come from the same distribution of values as the training data, then the model, on average, has to make the same percentage of errors as observed when the model was trained.
#Reproducibility
A model with similar properties can be built again from the same training data using the same algorithm and values of hyperparameters. No additional analysis, labeling, or coding is necessary to rebuild the model, only the compute power.
Principal Component Analysis 1