Open zhuqinghahaha opened 1 year ago
Unfortunately, it's been quite some time since I wrote the reproducibility file. However, there was a previous issue detailing certain problems with reproducibility and the user was able to match our results: https://github.com/allegro/allRank/issues/23.
Did you preprocess each fold separately using the provided script?
Thank you for your reply! I successfully reproduce the result after doing the features normalization.
LOSS | Self-attention | Self-attention | Self-attention | MLP | MLP | MLP |
---|---|---|---|---|---|---|
NDCGLoss 2++(Vali) | 0.52359 | 0.54353 | 0.5982 | 0.48959 | 0.51058 | 0.57051 |
LambdaRank(Vali) | 0.51851 | 0.53841 | 0.59356 | 0.48601 | 0.50747 | 0.56734 |
Although I followed the exact settings outlined in the reproducibility file, my experiments consistently yield inferior results compared to those reported in the paper. Any suggestions, recommendations, or additional details I might overlooked?
WEB30K—Result in paper
WEB30K— Reproduce