A common pitfall when training a BigQuery ML model is overfitting. Overfitting occurs when the model matches the training data too closely, causing it to perform poorly on new data. BigQuery ML supports two methods for preventing overfitting: early stopping and regularization.
To learn how to modify the options described below, see
CREATE MODEL statement.
Early stopping is the default option for overfitting prevention in
BigQuery ML. When early stopping is enabled, the
loss on the
holdout data is monitored during training, and
training is halted once the loss improvement in the latest iteration falls below
a threshold. Since the holdout data is not used during training, it is a good
estimate of the model's loss on new data. The
data_split_eval_fraction options control the behavior
of early stopping.
Regularization keeps the model weights from growing too large, preventing the model from matching the training data too closely. BigQuery ML supports two methods for controlling the size of the model weights: L1 regularization and L2 regularization.
By default, the values of
l2_reg are zero, which disables
regularization. On some datasets, setting positive values for
l2_reg will improve the trained model's performance on new data. The best
values for the regularization parameters are typically found through trial-and-
error, and it is common to experiment with values across several orders of
magnitude (for example, 0.01, 0.1, 1, 10, and 100).
Here is some general advice on using regularization:
If you are experimenting with the regularization parameters, try disabling early stopping so that the effect of regularization is clear.
If the number of features is large compared to the size of the training set, try large values for the regularization parameters. The risk of overfitting is greater when there are only a few observations per feature.
If you are concerned that many features may be irrelevant for predicting the label, try setting
l1_regto be larger than
l2_regand vice versa. There is theoretical evidence that L1 regularization works better when many features are irrelevant.
Another benefit of L1 regularization is that it tends to set many model weights to exactly zero, which is helpful for identifying the most relevant features and training a compact model.
- For an overview of BigQuery ML, see Introduction to BigQuery ML.
- To get started using BigQuery ML, see Create machine learning models in BigQuery ML.
- To learn more about working with models, see: