xgfs / deepwalk-c

DeepWalk implementation in C++

Hey! I am a Research Scientist at Google Research, NYC. I did my Ph.D. at the University of Bonn. I am interested in scalable, principled methods for analyzing graph data.

You can find more about my work on Google Scholar and personal website. Follow me on Twitter to get the latest updates.

deepwalk-c's Issues

Node (names | ids) in the output of the deepwalk-c?

Hey,

I've been playing with your implementation of deepwalk, and couldn't really wrap my head what is the format output of the binary file? If I read it into a numpy array how can I recover the ids of the nodes?

Do you have a suggestion how rewrite the binary output into an ASCII or UTF-8 format as below?

node_i emb^i_1 emb^i_2 ... emb^i_d
...
node_n emb^n_1 emb^n_2 ... emb^n_d

The performance problem

The performance of this implement is far from the original's.
Have you found whats the problem?

Segmentation fault

When constructing the HSM tree I face a segmentation fault. Here is the bcsr file that is the input. 1.9 GB download https://uni-bonn.sciebo.de/s/K9ECnxo1Z9sIHe6

Is this lib useful on directed graph?

Hi,
Thanks for your share.Have you ever try experiment on directed graph,I found that this worked good on undirected graph but have some porblem with directed graph.

Multiprocess utilization

It seems that adding more cores doesn't improve speed proportionally. Why is that?

when use GPU，can speed up this？

according to cuda, implement a version of GPU, can make significantly improving on performance?

How to read the output file by Python

Hi,

Thank you so much for the efficient code. I do not know how to read the binary output file by python. Can you provide some information about that?

Multiple memory leaks and memory corruption

With script and karate graph from pull request #2 I run

    convert-bcsr \
        --format mat \
        --matfile-variable-name adjmat \
        var/karate.mat var/karate.xgfs
    src/deepwalk -input var/karate.xgfs -output var/karate.emb

Then system sometimes dumps core. It observed the same effect if OpenMP is turned off.

how to evaluate of the correctness of the results?

we want to evaluate DeepWalk-c on the BlogCatalog dataset used in the DeepWalk paper.
using the scripts from https://github.com/phanein/deepwalk.
but we meet problem when construct the ouput vectors.
how to get the vertex number from the csr?

Where is the convert-bcsr tool?

Hi author:
Thanks for your share!
I have cline your project but I cannot find the convert-bcsr command.Is there anything else I need to do to find the command?

Source code decoupling

It would be better to restruct code and publish a kind of interface in order to everyone can use it as dependency. There are several issues.

1. Separate entry point in order to use deepwalk in different contexts.
2. Provide option to compile without OpenMP. Decoration of omp_get_thread_num() call with _OPENMP macro wil be nice.
3. Add some tests. For example, death test can help to detect wrong memory writes.
4. There is shared inner state that prevent usage as is in concurrent runtime.

The first two points are the most crucial for me since I would like to use it in side project.