To work with datasets, you need to create ContentWise and RL4RS class instances which filter and prepare the data. It takes time, so they can be dumped to pickle for further fast loading. Do something like this:
c = ContentWise('/home/USER/Downloads/data/ContentWiseImpressions/data/ContentWiseImpressions/CW10M-CSV/')
c.dump('cw.pkl')
r = RL4RS('/home/USER/Downloads/data/rl4rs-dataset/', 'rl4rs_dataset_b_sl.csv')
r.dump('rl4rs.pkl')
Warning: It takes ~15 minutes on a 2018's laptop to process and requires ~24GB RAM in total for both datasets.
To train and evaluate models, load datasets with ContentWise.load('dump_filename.pkl') or RL4RS.load('dump_filename.pkl').
You will get torch.utils.dataset. Its item is a dict with string keys and numpy array values. See ReciommendationDataset.__getitem__ for description. See Jupyter Notebooks in benchmark folder to get usage examples.
* benchmark/datasets -- main class with datasets wrapper, and some prepared **datasets.**
* benchmark/utils -- utility stuff, including models train/evaluate, batch collate function, and so on
* benchmark/evaluated_models -- models with computed metrics goes here to not mess in root directory
* benchmark/*|
|-- Jupyter notebook with different experiments
See RecommendationDataset class docstring and comments.
Also see comments in ContentWise and RL4RS, it might be helpful.
There is a class DummyData with few users and items. Maybe it can help you understand what's happening.
If you have any further questions or need additional assistance, feel free to ask!