For timeseries feature tables, an inner join is made at the creation of the feature table. For the other type of feature tables, a left join is made, so NaN values can show up in the training set. Can the inner join in create_training_set() method be implemented with a parameter?