A minor discrepancy between the code and the paper

[Rilkia]

Hi, team.
1.The formula (4) mentioned in the paper states that the representation obtained through the relational graph transformer is the weighted aggregation of neighbor representations. However, upon briefly understanding the mechanism of TransformerConv in the code, I found that its output is the node representation along with the weighted aggregated neighbor representations. I am unsure if my understanding is correct or if there might be an issue with it.
2.Regarding the construction of initial hidden vectors for the nodes in the relational graph, the description given in the paper is consistent with BotRGCN from 2021. That is, the initial hidden vector equals the tweet features, profile description features, 6 numerical features , and
11 categorical features. However, in the code, the number of numerical and categorical features doesn't match that in the paper. Could you explain the reasoning behind this inconsistency? Additionally, when conducting comparative experiments, how were the initial hidden vectors for BotRGCN's nodes obtained?
您好，感谢在百忙之中查看我的问题。
1.论文里面公式（4）提到，通过relational graph transformer，获取到的表示是邻居表示的加权聚合，但我在简要了解代码中TransformerConv机理时，发现它输出值为节点表示+加权聚合后的邻居表示，我不清楚是否是我的理解出现了问题
2.对于构建关系图节点的初始隐藏向量，论文里面给出的描述是与2021年 BotRGCN保持一致，即初始隐藏向量=推文特征+个人描述特征+数字特征（6个）+类型特征（11个），但代码中数字特征与类型特征的特征数与论文不匹配，这是出于什么考虑呢；在做对比实验时，是如何得到BotRGCN的节点初始隐藏向量的呢？

[Rilkia]

@BunsenFeng, sir please