首先，非常感谢您share SGL的代码，今天在看您代码的过程中，发现自己对sub_mat这个dict，并不是很理解，请问可以帮忙注释一下吗？非常感谢，这对一个初入门的选手太重要了。

    with tf.name_scope("input_data"):
            self.users = tf.placeholder(tf.int32, shape=(None,))
            self.pos_items = tf.placeholder(tf.int32, shape=(None,))
            self.neg_items = tf.placeholder(tf.int32, shape=(None,))

            self.sub_mat = {}
            if self.aug_type in [0, 1]:
                #0: Node Dropout; 1: Edge Dropout
                self.sub_mat['adj_values_sub1'] = tf.placeholder(tf.float32) 
                self.sub_mat['adj_indices_sub1'] = tf.placeholder(tf.int64)
                self.sub_mat['adj_shape_sub1'] = tf.placeholder(tf.int64)
                
                self.sub_mat['adj_values_sub2'] = tf.placeholder(tf.float32)
                self.sub_mat['adj_indices_sub2'] = tf.placeholder(tf.int64)
                self.sub_mat['adj_shape_sub2'] = tf.placeholder(tf.int64)
           else:
                #2: Random Walk
                for k in range(1, self.n_layers + 1):
                    self.sub_mat['adj_values_sub1%d' % k] = tf.placeholder(tf.float32, name='adj_values_sub1%d' % k)
                    self.sub_mat['adj_indices_sub1%d' % k] = tf.placeholder(tf.int64, name='adj_indices_sub1%d' % k)
                    self.sub_mat['adj_shape_sub1%d' % k] = tf.placeholder(tf.int64, name='adj_shape_sub1%d' % k)

                    self.sub_mat['adj_values_sub2%d' % k] = tf.placeholder(tf.float32, name='adj_values_sub2%d' % k)
                    self.sub_mat['adj_indices_sub2%d' % k] = tf.placeholder(tf.int64, name='adj_indices_sub2%d' % k)
                    self.sub_mat['adj_shape_sub2%d' % k] = tf.placeholder(tf.int64, name='adj_shape_sub2%d' % k)

首先感谢您的代码，然后亚马逊数据集中的amazon-book.train中的user字段和item字段代表的是交互过的意思么？那么为什么amazon-book.train中的user，item是连续的，而amazon-book.test中的user，item是不连续的以及这字段的实际意义代表的是交互过的意思么？谢谢

Hello

I reported this error while running your code. Do I need to add a validation mechanism here, or is there a problem with my super-parameter adjustment? Hope to get your guidance, thank you!

Hello, Have you ever tried to add projection (e.g., MLP) like SimCLR? I have add such projections, but it doesn't work. The performance is not good and the convergence is relatively slow. Are there any suggestions for projection function in SGL?

https://github.com/wujcan/SGL/blob/fcfb31f3af9a41fc2f6bc81932733952b91bbc90/model/general_recommender/SGL.py#L454-L460

对于每个batch累计的各类loss，为什么打印的时候，除以的是 data_iter.num_trainings （数据集的大小）呢？

我看LightGCN里面除以的是dataset的batch的个数

Hello, I am working on SGL recently. Thank you for offering this project. In sec 4.3.3, it said that

we contaminate the training set by adding a certain proportion of adversarial examples (i.e., 5%, 10%, 15%, 20% negative user-item interactions), while keeping the testing set unchanged. Figure 6 shows the results on Yelp2018 and Amazon- Book datasets.

May I ask how to get the adversarial examples? Is it just to add random links?

I use amazon-book (keep the rating > 3), and add the interaction which score =1/2 as adversarial examples. But I don't find such performance degradation.

Something strange happened when I ran the codes. I ran it with yelp2018 data set, but each epoch took about 15 minutes, while my GPU did not have a lot of work. Is it normal for the training speed to be slow? Or is there something that I have missed?
PS: I've used cython and cpp implements for evaluate.

您好！非常感谢您分享的代码 我注意到您在论文里说您的方法与图的模型无关 请问您做过自监督辅助任务损失 和 推荐系统的监督学习任务的损失 比例的相关实验研究吗？自监督辅助任务损失 和 推荐系统的监督学习任务的损失 在数值上以什么样的比例能取得较好的性能提升 或许您也可以提供一些参数优化的相关思路 期待您的回复！

Hello, I have changed my root path ,but another error occurred.

This is it:

Traceback (most recent call last): File "/(#my root path#)/SGL/main.py", line 22, in conf = Configurator(root_folder + "NeuRec.properties", default_section="hyperparameters") File "/(#my root path#)/SGL/util/configurator.py", line 68, in init self.alg_arg = self._read_config_file(arg_file) File "/(#my root path#)/SGL/util/configurator.py", line 88, in _read_config_file raise ValueError("'%s' is empty!" % filename) ValueError: './conf/MF.properties' is empty!

你好，最近在看你论文，同时尝试运行了下代码，发现有报错，内容如下： Traceback (most recent call last): File "main.py", line 23, in seed = conf["seed"] File "/self-sup/SGL/util/configurator.py", line 128, in getitem raise KeyError("There are not the parameter named '%s'" % item) KeyError: "There are not the parameter named 'seed'" 请问下如何解决这个问题，谢谢。

您好，最近在研究您这篇论文中的代码，在最后predict和evaluate的时候用的是代码中的ua_embeddings，和ia_embeddings（个人对代码的理解），而不是两个subgraph的embedding，所以想请教一下最终结果和两个subgraph的embedding是否有联系，如果您是通过两个subgraph的embedding得出的ua_embeddings，和ia_embeddings，想请教这步的具体过程在代码中哪个方位，或者大体思路，谢谢您

Thanks for your great work. I saw proofs of equations 12, 13, 14 in your supplementary material. But I still don't know the detailed derivation process of Equation 15, that is: why the L2 norm of c(v) is proportional to the following term:

Hi, thanks for your great work.

I am confused about robustness to noisy interactions in this paper.

Towards this end, we contaminate the training set by adding a certain proportion of adversarial examples (i.e., 5%, 10%, 15%, 20% negative user-item interactions), while keeping the testing set unchanged.

I tried to sample from interations which don't show in train.txt and test.txt. But I didn't find too many differences between LightGCN and SGL. I wonder if it is proper to generate adversarial examples in this way.

Looking forward to your reply.

是关于自监督损失的复杂度方面的，文中给出的推导过程是 在一个批次的计算中，分子部分:O（Bd） + 分母部分O（BMd） -> 一个批次的总复杂度为 O（|E|d(2+|V|)）

关于这个推导不是很理解，目前暂时还没有看源码，所以也不清楚具体的实现是怎么样，我自己的推导过程是 one epoch: user = bd+bmd=b(m+1)d item = bd+bid+b(i+1)d user+item = b(m+2+i)d -> O（bd(2+|V|)） 区别就在 bd(2+|V|) 还是 |E|d(2+|V|) 因为我本人刚入门，可能问题有点蠢，还望您能够解答，文章的其他部分还是能读下来的

您好，最近在尝试你的方法，但是代码有如下报错： (tensor) algroup@algroup:~/lxj/SGL-main$ python main.py --recommender=SGL seed= 2021 WARNING:tensorflow:From main.py:27: The name tf.set_random_seed is deprecated. Please use tf.compat.v1.set_random_seed instead.

split and save data... 2022-03-07 10:08:39.052: amazon-book_given_u0_i0 2022-03-07 10:08:39.053: Dataset name: amazon-book The number of users: 52643 The number of items: 91599 The number of ratings: 2984108 Average actions of users: 56.69 Average actions of items: 32.58 The sparsity of the dataset: 99.938115% Item degree grouping... Traceback (most recent call last): File "main.py", line 33, in dataset = Dataset(conf) File "/home/algroup/lxj/SGL-main/data/dataset.py", line 37, in init self._group_item_by_popularity() File "/home/algroup/lxj/SGL-main/data/dataset.py", line 262, in _group_item_by_popularity self.item_group[i] = i_degree[self.item_group_idx == i] AttributeError: 'Dataset' object has no attribute 'item_group'