MetaDataset

MetaDataset is the key component for splitting dataset by task.

Overall, we extend Dataset to MetaDataset.

The extended modification can be listed briefly as following:

• [Override] self.bulid(): Add task keyword for group_by.

• [Add] self.split_by_ratio_meta(ratios, group_by): Split method by task.

self.build()

In metaDataset, we add a new eval_args.group_by keyword task in it, which can split the dataset by user clusters.

If we set the keywords task and RS, it will call self.split_by_ratio_meta(), which we design it for the split method.

self.split_by_ratio_meta(ratios, group_by)

This function is used to split dataset by non-intersection users clusters, which means the interactions in the output datasets (eg. training dataset, valid dataset and test dataset) have non-intersection user clusters.

Split by non-intersection users is significant for user cold start task in meta learning recommendation.

ratios

The split ratios of user clusters.

group_by

Commonly task, and only by task can this function be called.

[return] next_ds

[dataset_1, dataset_2 ... dataset_n]