Description

Currently our unit tests are disorganized and each test creates example StellarGraph graphs in different or similar ways with no sharing of this code.

This issue is to improve the unit tests by making functions to create example graphs available to all unit tests by, for example, making them pytest fixtures at the top level of the tests (see https://docs.pytest.org/en/latest/

我觉得这里的else：row.append(i + vocab_size)应该改为else：row.append(i + train_size+vocab_size)，对吗？
这段代码是为doc节点和word节点生成稀疏邻接矩阵的代码，邻接矩阵的大小为train_size + vocab_size + test_size， 当doc文本序号i大于train_size时，剩下的不就是test_size大小的文本与单词建立连接吗？test_size在邻接矩阵之前不是有train_size+vocab_size，所以此时是不是文本从train_size+vocab_size开始一一与词建立连接？

以下为源码：

for i in range(len(shuffle_doc_words_list)):
    doc_words = shuffle_doc_words_

nowadays, docs with rarely interpretation is difficult to understand the algorithm,
Such as bayes rule sets and other Underdogs have little references.
If can provide a common introduce in interface level may be good for programmers
who have little information of algorithm to start

I noticed in example code, some argument are inconsistent in the commented parts, such as lines in examples/molnet/train_molnet.py:
https://github.com/chainer/chainer-chemistry/blob/56e83dedb5de9dc9eb08ebf292be9ba76a4883ba/examples/molnet/train_molnet.py#L267-L274

args.gpu and concat_mols are not consistent with other codes and will throw errors. They should be modified to device and

The README.md describes what KGCN is, but it does not describe how it will be beneficial for users.

We should have a use-case section describing the kind of problems in which KGCN makes sense as a solution.

I viewed the whole code and found that the code only use toy dummy data to train model. So I don't really understand how you use those data to train GCN model. Can you supply the code or instructions about how to use real-world data to train model?