junkangwu/adap_tau

[WWW 2023] Official code of "Adap-$\tau$: Adaptively Modulating Embedding Magnitude for Recommendation"

adap_tau's Issues

How to get equation(7) in the paper

Sorry to be a bother! I think it's excellent work but I'm confused about how to get equation(7) in the paper, it appears that I cannot obtain this result if I adhere to the derivative calculation method of softmax cross-entropy loss

About dataset

Hello author, I don't see the data set of Movielens in the source code, could you please provide the data set of Movielens? thank you.

about the tau in traing process

Thanks for the excellent work by the authors. I'm confused about the values of tau when I run the code by:

bash Adap_tau_novel.sh amazon-book 1e-3 1e-7 3 1024 2048 nopdrop 1.0 1.0 uniform_gpu 0 100 cosine mf weight_mean

here's an example in a epoch:

2023-05-05 16:20:39,454 - [INFO] - current w_0 is 9.65932889229748
2023-05-05 16:22:47,693 - [INFO] - E:6|TAU:7.724 17.2, train_time: 126.4, VALID_time: 1.813, loss: 3.757e+03, emb_loss:1.48e-05, best_valid(recall@20): 0.04892
valid N@10: 0.03014, R@10: 0.02926, P@10: 0.02303
valid N@20: 0.03826, R@20: 0.04941, P@20: 0.01996
valid N@50: 0.05434, R@50: 0.09246, P@50: 0.01561

I know they are the min and max of tau, Are they the tau used in SSM loss? If so, I'm wondering why the values are so large since the common range of tau is [0,1].