[Question] Split train and test data by timestep about recbole

[hyp1231]

This function can be abstracted as splitting the original interaction records according to values of the certain feature (split_by_value), which should be another choice besides split_by_ratio. This part has been considered but not been implemented yet. PR about this function is highly welcomed. Please see

RecBole/recbole/data/dataset/dataset.py

Lines 1356 to 1364 in 948d9d6

	split_args = eval_setting.split_args
	if split_args['strategy'] == 'by_ratio':
	datasets = self.split_by_ratio(split_args['ratios'], group_by=group_field)
	elif split_args['strategy'] == 'by_value':
	raise NotImplementedError()
	elif split_args['strategy'] == 'loo':
	datasets = self.leave_one_out(group_by=group_field, leave_one_num=split_args['leave_one_num'])
	else:
	datasets = self

for details.

Besides the above discussion, there is indeed another way to split train and test data by certain timestamps. You can pre-split the .inter into several new atomic files by timestamp, and use the config benchmark_filename to load the pre-split datasets. Please refer to #586 (comment) for details.

[johnny12150]

Ok, I will take a look at it.

benchmark_filename: ['part1', 'part2', 'part3']

Is it correct that part1 will be automatically loaded as training data?
Edit
Yes.