Available in: GBM, DRF, Deep Learning, Naïve-Bayes, AutoML
During model training, you might find that the majority of your data belongs in a single class. For example, consider a binary classification model that has 100 rows, with 80 rows labeled as class 1 and the remaining 20 rows labeled as class 2. This is a common scenario, given that machine learning attempts to predict class 1 with the highest accuracy. It can also be an example of an imbalanced dataset, in this case, with a ratio of 4:1.
balance_classes option can be used to balance the class distribution. When enabled, H2O will either undersample the majority classes or oversample the minority classes. Note that the resulting model will also correct the final probabilities (“undo the sampling”) using a monotonic transform, so the predicted probabilities of the first model will differ from a second model. However, because AUC only cares about ordering, it won’t be affected.
If this option is enabled, then you can also specify a value for the
max_after_balance_size options to control the sampling.
class_sampling_factorstakes a list of numbers which would be the sampling rate for each class. A value of
1would not change the sample rate for a class, but setting it to
0.5would reduce its sampling by half, and
2would double its sample rate.
Alternatively, you can utilize
max_after_balance_sizewhich is the max relative size your training data can be grown. By default, it is
5: this will oversample the data to rebalance the training data. The max it can grow to is 5x larger than your original data, hence, the value of 5. If you have many rows and prefer to under-sample the majority class, you can set
max_after_balance_sizeto a value of less than
This option is disabled by default.
This option only applies to classification problems.
Enabling this option can increase the size of the data frame.
Refer to the following link for more information about balance classes: https://gking.harvard.edu/files/0s.pdf.