cszhangzhen / anrl Goto Github PK

Hi, I have two questions concerning about ANRL:

1. Can ANRL run the model with customized continuous attributes? (no obvious statement)
2. Can ANRL run the model without label? If it can make it, how can I do some modifications to this code?
3. Is this method capable of handling weighted edges? If it can make it, do I need to make further modifications on your code? (I indeed found your answer toward other person, but you annotated that code with pass)

continuous attributes example:

0 1100255.0 9.1 29.5261 248515.0 0.2351 0 0 0 0 1 0 0 0 0 0 0 0
1 228151.0 5.64 41.8182 92073.0 0.4239 0 0 0 0 0 1 0 0 0 0 0 0
2 131061.9 0.5345 34.2327 32476.5 0.2621 0 1 0 0 0 0 0 0 0 0 0 0
3 222647.0 7.21 -6.3848 43273.0 0.1299 0 0 0 0 0 1 0 0 0 0 0 0

Look forward to your reply!

I hope that you can share the other data sets of the test. The data set on the network cannot meet the experimental conditions. Your help will be very useful to me. Thank you.

Excuse me, I wonder how to apply ANRL to Pubmed dataset?
I changed the read_feature function to read Pubmed feature matrix like:
features = sio.loadmat(inputFileName)
features = features['feature']
I adjusted the parameter in tf.app,flags and changed the struct parameter in config.py to [None, 200 ,None].
However the result is confusing, please tell me what other operations should I do? Thanks a lot.

I have noticed that the code for contructing targeted neighbours while the graph is weightd has been annotated,
I think maybe there exists a bug as follows:
temp = X[node] , which is a shallow copy of X, so X will be changed after the temp is changed
we should use a deep copy, temp = np.array(X[node])

Hi,
After I read the paper, I suppose that you enable model to learn topology and attributes information at the same time by adding three loss functions together. However, I noticed that this model may not infer the topology structure well. Is it possible to improve it by adding one hot encoding value for model input(The model just use attributes as input)?

I am not quite clear about this part of your code
def encoder(X):
for i in range(self.layers - 1):
name_W = 'encoder_W_' + str(i)
name_b = 'encoder_b_' + str(i)
X = tf.nn.tanh(tf.matmul(X, self.W[name_W]) + self.b[name_b])
return X

Why bother using a for loop here, considering you only get the last value of this for loop as function return.

您好，我是一位网络表示学习初学者，请问您提出的这个模型是无监督的还是有监督的？因为我看到链路预测的无监督任务以及节点分类的有监督任务，并且您的代码里面用到了标签信息。请问这个模型是有监督无监督都能做吗，无监督版本该如何实现呢

Hello, your thoughts make me refreshed. I am very curious about reading the article. What is the attribute information you have proposed? After the research of the code, is it the adjacency matrix corresponding to the network?

如果用在异构图上，图中会有多种节点，每种节点的特征维度可能不一样。想问作者有没有解决方法？

作者您好！您能不能将下面这个格式的数据转化成citeseer.edgelist citeseer.feature的代码开源出来啊？
ind.cora.allx
ind.cora.ally
ind.cora.graph
ind,coratest.index
ind,coratx
ind,cora.ty
ind,cora,x
ind.cora.y

Hi,
If I input edgelist like this:
(weighted and directed)
0 456 3.5229
0 78 3.3552
0 1022 1.6327
0 989 1.3817
0 56 1.1477
……………………
And feature list like this: (I think 0 means ignoring the feature)
1567 5
0.000000000000000000e+00 0.000000000000000000e+00 0.000000000000000000e+00 0.000000000000000000e+00 0.000000000000000000e+00
0.000000000000000000e+00 0.000000000000000000e+00 0.000000000000000000e+00 0.000000000000000000e+00 0.000000000000000000e+00
0.000000000000000000e+00 0.000000000000000000e+00 0.000000000000000000e+00 0.000000000000000000e+00 0.000000000000000000e+00
……………………
Then I get the same result for every node:
0,0.27159807,-0.5618622,0.389931
1,0.27159807,-0.5618622,0.389931
……
So why the different weights seem have no influence on the embedding result?
Look forward to your reply!

您好！ANRL非常简洁高效！请问对于pubmed数据集，超参应该如何设置，对于这个数据集，我得到的结果一直很低，请问针对这个数据集，除了自动编码器的隐含层，还应该如何设置其他超参来得到理想的效果！谢谢！期待您的回复！