Keep item-id-mapping the same between multiple dataset filterings

[MatthiasEg]

Hi,

I want to evaluate my models' performance on various "user_inter_num_interval"'s to check for the cold-start problem. However, when I filter the datasets users given some interval and apply the model I found that the models performance is terrible (compared to how it performed on the unfiltered dataset). It's impossible that the model performs so bad, just by removing 1% of sequences.

Given that, I now found that the remapping (remap-id) of item-id's is inconsistent when different filters are applied (when the interactions and users change), hence the models item-embeddings do not align anymore with the item-id's.

How could such an issue be solved? Can I somehow disable the remap-id? How else would I test for the cold-start problem, while ensuring that the RecBole-internal item-id-mapping stays exactly the same?

Thanks in advance for your help!

[BoXiaohe]

Thanks for your attention to RecBole! As for your problem, maybe you can try processing the dataset first before fit the data into recbole models. Hope this could help you!