The sequential recommendation method implemented in this library is trained and tested on datasets such as Beauty and Sports, and the results obtained are significantly higher than those reported in the original paper. Could you please provide an explanation?

[SONGLEI-arch]

[yuh-yang]

Thank you for raising this question.

Could you please clarify which specific models and which original papers you are comparing against? Having the exact references would help us investigate the discrepancy more accurately.

In general, a common reason for this kind of misalignment lies in differences in evaluation protocols. Many prior works report results using sampled evaluation, where the target item is ranked among a small set of negatively sampled items (e.g., N=100 or N=200). Such sampled metrics often inflate performance and are not directly comparable across different implementations.

In contrast, SSLRec adopts a unified full-rank (all-item) evaluation setting, which avoids negative sampling and provides a more unbiased and consistent measurement of model performance. As a result, performance numbers obtained under full-rank evaluation can differ significantly from those reported in papers that use sampled metrics, even when the same model is implemented.

Once we know the exact models and papers you are referring to, we would be happy to provide a more detailed, case-specific explanation.