woeee / fmlp-rec Goto Github PK

您好！
我们在对模型采用full sort参数时，尝试了一些超参数，得到的结果都与您在论文附录中提供的数据相差较大。请问您能给出当采用full sort时，FMLP模型在几个数据集上的超参吗？
期待您的回复。

Hi, thanks for your great work!
But I have a question. I noticed that when using the parameter full sort, you use full items in validation and testing. Have you tried using the full sort during training as well, and how does it compare to that?
Looking forward to your reply！

Hi! Thanks for such solid work. I have some kind questions about details in implementation.

1. There are many self-implemented functions, e.g., LayerNorm, gelu. Why did not use the integrated component in torch? (I have tested the F.layernorm would give the same results)
2. In full-sort setting, is the seed set to 42 or else? What about other hyperparameter settings for other models?
3. In FMLPRecDataset, the last three items are used for training target, eval label and test label. What if the total sequence is no longer than 3? For example, let self.user_seq = [[1,2,3]], the parsed input_ids would be [[]] and further padded as 50 zeros, and the trainer seeks to utilize [[0,0,...,0]] to forecast [[1]]. Maybe exploiting such all-zero sequences to estimate targets would introduce extra noise?

Hi there!

Very glad to see your interesting work! It's novel and effective.

Sadly, I have an issue about the 'full_sort' setting for evaluation. The code in this line shows that you have filtered the users' historically interacted items in the full item table and then calculate the metrics based on the filtered items. What confused me is that, in the e-commerce scenario, many users will have repeated consumption behaviors. The next interacrted item is prone to be an item interacted before. And many of your experiments are conducted on the e-commerce datasets, e.g., the Amazon dataset. So my question is that is it reasonable to filter out the historically interacted items before evaluation in the full-sort setting. Morever, could you please provide any references for the full-item setting? I don't see references in the paper.

Anyhow, the experiments in full-sort setting do not badly affect the robustness of your paper. I just want to know why you have filltered out the interacted items.

Thanks in advance!

您好~这是一份很健壮的基线。这段时间我们对您的工作进行了深入研究和改进，目前在其他数据上取得了一定的提升；但是在session-based recommendation上和论文中的报道结果还有较大差距。

我们观察到您可能为每个数据集分别配置了不同的超参数（例如在full-sort评测协议下）。可以请教一下您四个session-based数据集上，fmlp的超参数设置和随机种子设置吗？如果是多个随机种子平均的话，分别是哪些随机种子呢？