Hi,
I'm a master degree student and I've read your work.
I'm trying to reproduce your experiments with Last.fm DB, but the README is very synthetic.
I use this configuration for preprocessing:
python preprocess.py --max_session_len 20 --suffix _session20_occur10-5 --op all
and this configuration for testing:
python main.py --train_file clean-train_session20_occur10-5 --test_file clean-test_ses--batch_size 64 --name ripr_paper

Is it correct? I can't understand how to get the dataset written in Table1 with 767 user and 107606 items. With these parameters I got a larger dataset (probably more sparse) and in testing results were:
Recall@5: 0.2153 - MRR@5: 0.1750
Recall@20: 0.2742 - MRR@20: 0.1810

The output of preprocessing is this:

• Total origin sessions 656061
• Total cut sessions: 1122490
• Event per origin sessions 23.4229469515792
• Event per cut sessions 13.689994565653146
• Total events: 15366882
• Average sessions length: 23.4229469515792
• Sessions per user: 674.9598765432099
  Third filter (remove item not exist in the train set):
• Num users: 972
• Num items: 254709
• clean-train_session20_occur10-5new: 14353289 events
• clean-test_session20_occur10-5new: 717352 events
• clean-dev_session20_occur10-5new: 735971 events

very different of Table 1 page 8 of the paper.

Can you help me to reproduce the dataset preprocessing and the PostAdapt NSAR like the paper?
Thank you very much.

EDIT: I can't understand the difference of this two parameters:

parser.add_argument('--max_valid_seq_len', type=int, default=500,
                    help='The maximum length allow per session')
parser.add_argument('--max_session_len', type=int, default=10)

and how to set the overlap of 5 events (page 8 of paper)