Hi, thanks for sharing the source code for reproducing your results. However, I am confused with mapping the hyper-parameters between your code and paper:

1. lambda is placed on L_C for constraint loss in paper (Eq. 12, 18) for U-I relationships. But, in code, lambda is used for L_2 loss?

lambda's setting is within [0.2, 0.4, 0.6, 0.8, 1.0, 1.2, 1.4] in paper, but in code we have gamma=1e-4?

2. gamma for I-I relationships in paper (Eq. 18) and is inline with the code. And its setting [0.1, 0.5, 1.0, 1.5, 2.0, 2.5, 3,3.5] in paper. However, in code, we observe gamma with 5e-4 or 1e-3 (ML-1M). Should we follow the paper for its settings?

3. w1, ..., w4 are used for generating omega coefficient, but you have mentioned they are related with lambda in previous issue 12?

thank you for your paper I read LR-GCCF, NGCF papers before and the LR-GCCF result was better than NGCF how did you test these 2 models? and how did you find the best hyperparamters of two models? could you tell me? I want to test models by myself.

Thanks for your interesting work! I found the ml-1m dataset provided in this repo may be inconsistent with that reported in the original paper, see the data statistic table in 4.1. Only 895699 interactions in your dataset but 995154 are reported. Could you please update the dataset or explain it? I also wonder how you process the dataset (e.g., filter users or items, split train/validation/test datasets)? Looking forward to reply!

Hi! Thanks for the great work. There are two convergence states in eq.9 and eq.16 in your paper, may I ask the process of deriving \beta_{u,i} and \omega_{i,j}?

Hi, Recently I came across your paper ( a nice work), your paper explained eq. 10 with "We normalized the embeddings to unit vectors and then maximized the dot product of the both terms". However, I can not find "normalization" operation in your codes. May I ask you about this part? Thanks a lot

作者您好，感谢您的工作！ 想请教一下在原文的3.3节中对时间复杂度的推导里得到的结果 O（（K+R+1）|A+|(d^2+1））的具体推导过程？主要的问题是在d^2上，这里我预先自己推的结果是(d+1)(没有平方)，粗浅的理解是算损失只需要计算d维度的乘法，同时针对每个交互样本计算一次对应系数。 能否请教一下具体的过程呢？

Hi! What is the relationship between the hyper-parameter lambda described in your paper and the variables w1, w2, w3, w4 in the code? The paper reported that "We tune lambda in [0.2, 0.4, ..., 1.2, 1.4], ...". However, in many dataset_name_config.ini files, some w1,...,w4 are set to 1e-8 or 1e-7. For example, in amazon_config.ini: #L = -(w1 + w2*\beta)) * log(sigmoid(e_u e_i)) - \sum_{N-} (w3 + w4*\beta) * log(sigmoid(-e_u e_i')) w1=1e-8 w2=1 w3=1 w4=1e-8

Thank you!

Hi authors, I noted that you tested your UltraGCN model on another three datasets: MovieLens-1M, Amazon-CDs, Amazon-Electronics. Could you please provide those datasets so that we can better follow your work? Much appreciated.

In your paper, we see that L_I is beneficial. Because L_I put constraint on the item-item graph. Similar to L_I, the constraint loss L_U on the user-user graph should also be important. Have you tried the effect of L_U?

Compute weight in real time, without storing constraint_mat.

Only tested on ml-1m dataset, single epoch time improved from 20 seconds to 13 seconds on my own device.

No significant change in the evaluation metrics.