this code is not consistent with paper

[saladcat]

        for k in range(0, self.n_layers):
            if self.layer_fun == 'gcf':
                ego_embeddings_s = one_graph_layer_gcf(A_fold_hat_s, ego_embeddings_s, weights_s, k)
                ego_embeddings_t = one_graph_layer_gcf(A_fold_hat_t, ego_embeddings_t, weights_t, k)
            if k >= self.n_layers - self.n_interaction and self.n_interaction > 0:
                if self.fuse_type_in == 'la2add':
                    ego_embeddings_s, ego_embeddings_t = self.s_t_la2add_layer(ego_embeddings_s, ego_embeddings_t,
                                                                               self.lambda_s, self.lambda_t,
                                                                               self.domain_laplace)

            norm_embeddings_s = tf.math.l2_normalize(ego_embeddings_s, axis=1)
            norm_embeddings_t = tf.math.l2_normalize(ego_embeddings_t, axis=1)

            if self.connect_way == 'concat':
                all_embeddings_s += [norm_embeddings_s]
                all_embeddings_t += [norm_embeddings_t]
            elif self.connect_way == 'mean':
                all_embeddings_s += norm_embeddings_s
                all_embeddings_t += norm_embeddings_t
        if self.connect_way == 'concat':
            all_embeddings_s = tf.concat(all_embeddings_s, 1)
            all_embeddings_t = tf.concat(all_embeddings_t, 1)
        elif self.connect_way == 'mean':
            all_embeddings_s = all_embeddings_s / (self.n_layers + 1)
            all_embeddings_t = all_embeddings_t / (self.n_layers + 1)

where

            norm_embeddings_s = tf.math.l2_normalize(ego_embeddings_s, axis=1)
            norm_embeddings_t = tf.math.l2_normalize(ego_embeddings_t, axis=1)
            if self.connect_way == 'concat':
                all_embeddings_s += [norm_embeddings_s]
                all_embeddings_t += [norm_embeddings_t]
            elif self.connect_way == 'mean':
                all_embeddings_s += norm_embeddings_s
                all_embeddings_t += norm_embeddings_t

why you add norm_embedding to all_embeddings_s?
this code is not consistent with paper !!!

[sunshinelium]

首先非常感谢您对我们工作的兴趣。代码中这里是分为了两种层融合策略进行的实验，论文中考虑的是self.connect_way=='concat'的情况。

[saladcat]

norm_embeddings_s = tf.math.l2_normalize(ego_embeddings_s, axis=1)
抱歉可能我没表达清楚意思，请问这里为什么要做 L2呢？

[sunshinelium]

哦哦，这个地方是参考的ngcf的做法，加入了正则化项，在论文的公式17中写了。

…

---Original--- From: "ShiyangZhu"<notifications@github.com&gt; Date: 2021/3/5 15:57:29 To: "sunshinelium/Bi-TGCF"<Bi-TGCF@noreply.github.com&gt;; Cc: "sunshinelium"<sunshinelmj@qq.com&gt;;"Comment"<comment@noreply.github.com&gt;; Subject: Re: [sunshinelium/Bi-TGCF] this code is not consistent with paper (#3) norm_embeddings_s = tf.math.l2_normalize(ego_embeddings_s, axis=1) 抱歉可能我没表达清楚意思，请问这里为什么要做 L2呢？ — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[saladcat]

可是这个l2正则项为什么要在后面concat到一起呢？而且这里是多层GCN卷积的部分，正则项也不应该在这里出现呀。

[sunshinelium]

这个地方应该不是正则化项，只是通过对特征进行正则化来降低了特征的值，避免特征值大可能带来的梯度问题吧，这里我完全采用的NGCF代码中的处理方式，我的理解是这样，或许你可以问下NGCF的作者？

…

---Original--- From: "ShiyangZhu"<notifications@github.com&gt; Date: 2021/3/5 16:05:19 To: "sunshinelium/Bi-TGCF"<Bi-TGCF@noreply.github.com&gt;; Cc: "sunshinelium"<sunshinelmj@qq.com&gt;;"Comment"<comment@noreply.github.com&gt;; Subject: Re: [sunshinelium/Bi-TGCF] this code is not consistent with paper (#3) 可是这个l2正则项为什么要在后面concat到一起呢？而且这里是多层GCN卷积的部分，正则项也不应该在这里出现呀。 — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[saladcat]

好的了解到，感谢您的回复。