When using the sample_feature_values function to select cruises to use for final validation, it was thought a good idea to sample a fraction from each year, to ensure we good representation in the train and test sets from each year. This seems to cause dramaticx reduction inperformance. I think this means there is a bug, because there is a lot of custom code beyond standard pandas and scikit learn, so it seems likely that is causing the spplit to happen incorrectly and so poor results.
When using the
sample_feature_values
function to select cruises to use for final validation, it was thought a good idea to sample a fraction from each year, to ensure we good representation in the train and test sets from each year. This seems to cause dramaticx reduction inperformance. I think this means there is a bug, because there is a lot of custom code beyond standard pandas and scikit learn, so it seems likely that is causing the spplit to happen incorrectly and so poor results.The line is experiment.py
ensemble_unseen_cruise_numbers = self.xbt_labelled.sample_feature_values(self.unseen_feature, fraction=self.ens_unseen_fraction, split_feature='year')
currently we have removed the
split_feature
argument until we can fix the bug.