apple2373/metapath2vec

metapath2vec with tensorflow

This repo contains an implementation of metapath2vec using tensorflow.

main reference appeared at KDD 2017: metapath2vec:

Please use the author's implementation for formal experiments!!

I just wrote this for myself in order to learn how the algorithm works. I haven't tested on a big network or even checked if I can reproduce the reported performance, so be careful when you use it.... (I mean there might be a bug :) ). The author 's implementation is available here: https://ericdongyx.github.io/metapath2vec/m2v.html.

Requirements

I recommend you to install Anaconda and then tensorflow.

• tneosorflow
• and some other libraries...

How to use.

See help for the information. It should be self-contained.

python main.py --help

You need to provide two files: a text file that has node type information, and text files that has paths generated by random walks guided by a meta path. See data/test_data to find sample txts. Note that you have to generate meta-path guided random walks by yourself.

How to train.

learn embeddings using the random walks

python main.py --walks ./data/test_data/random_walks.txt --types ./data/test_data/node_type_mapings.txt --log ./log --negative-samples 5 --window 1 --epochs 100 --care-type 0
python main.py --walks ./data/test_data/random_walks.txt --types ./data/test_data/node_type_mapings.txt --log ./log --negative-samples 1 --window 1 --epochs 100 --care-type 1
tensorboard --logdir=./log/

how to load the learned embeddings

import numpy as np
import json
index2nodeid = json.load(open("./log/index2nodeid.json"))
index2nodeid = {int(k):v for k,v in index2nodeid.items()}
nodeid2index = {v:int(k) for k,v in index2nodeid.items()}
node_embeddings = np.load("./log/node_embeddings.npz")['arr_0']

#node embeddings of "yi"
node_embeddings[nodeid2index["yi"]]

To do list

• Make the batch size more than 1.

metapath2vec's Issues

meta-path generates random walk?

Question:
Hi:
How is your random walk corpus generated?Have you ever written code?

                                                                                                                                                         Thanks!

which data file in the paper corresponds ’node_type_mapping_txt‘ file in this codes？

On the author's website，I can not find the corresponding text files which has node types

Have you used this code to verify the experiment in the paper? How is the result compared to the paper?

Have you used this code to verify the experiment in the paper? How is the result compared to the paper?

How to train the model using batch_size?

Looks like I need to modify “get_negative_samples” function in dataset.py. However, i meet some bugs that cant be fixed...

how long does it need to train in dbis dataset and does it need gpu to speed up?what is the version of your tensorflow?thank you very much!

How the get meta path?

Thanks for provide this tf implementation. I'm confused about the meta path. If I have a heterogeneous Network with only two kinds of nodes, denote A and B. Is there only two kinds meta path A-B-A and B-A-B?

It seems like no implementation of metapath-based random walks.

It seems like no implementation of metapath-based random walks. Any help for finding it? or is it quit easy to implement it ?

pros with node embeddings

Hi, thanks for your amazing work. I wonder what's the meaning of the node embedding and index2node, it's the relation computed by the model?

metapath strategy: strictly or non strictly?

for v ∈ V do
MP = MetaPathRandomWalk(G, P, v, l) ;
X = HeterogeneousSkipGram(X, k, MP) ;
end
if the type of v is not equal to the first of your input metapath P, then you will skip it and use the next node v or choose the neighborhood node of v in its appropriate type in P. For example, the metapath is "author-conf-author" and node v is a conf, i wonder should i choose one author node of conf v according the partial path "conf-author" of the metapath P? or we skip the node v until we get an author node to match the metapath "author-conf-author" strictly ?
I appreciate your answer on it, Thanks!

where can I get the implementation of the first author of the paper

where can I get the implementation of the first author of the paper ？I can‘t find it in github。
thank you very much if you can give me a link。