[DOC] Change criteo example notebook performance metric(s) and validate model

[rdipietro]

Right now, examples/criteo-example.ipynb doesn't have a meaningful measure of performance:

• the performance metric is accuracy, which isn't meaningful because of the extreme class imbalance (about 97 negatives for every 3 positives)

Also the quality of the current model is possibly no better than a simple majority-vote classifier (e.g., def model(x): return 0):

• Performance outputs are currently not saved as part of the notebook
• If we run the notebook using day 0 for training and day 1 for testing, we achieve 96.5% accuracy
• The accuracy of def model(x): return 0 is (fraction of examples with label 0), which is about 97%.

So it would be beneficial to

• Include a more meaningful metric, e.g. AUC as done in the Joy of Cooking DLRM example
  • Implementation note: AUC can't be computed on a per-batch basis. In the context of fast.ai, see the AUROC class
• Validate the model using the full dataset once this metric is included