chang861224/hpe_struc

Struc2vec on Heterogeneous Preference Embedding

Preprocess

Step by step

Before you compile this project, you have to get the submodules first.

git submodule init
git submodule update

After getting the submodules, you have to enter the repository src/struc2vec. Because here is another submodule word2vec for submodule struc2vec, you have to repeat the commands above to get the codes of this submodule.

Then, you have to enter the repository src/word2vec in struc2vec and compile this project with following commands.

mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ../
make

After these procedures, you can go back to the root of this project and compile it.

make

Finally, you can use this project!!

Shell Script for Compiling Whole Project

If you think that the steps above are complicated, you can run the shell script file submodule_preprocessing.sh so that you can easily compile the whole project.

sh submodule_preprocessing.sh

Usage

It is almost the same as HPE in smore project. The most different is that there is another option -use_structure, which determines whether you should use the structural information or not. (Default is false.) So the options description are as follows:

Options Description:
    -train <string>
        Train the Network data
    -save <string>
        Save the representation data
    -dimensions <int>
        Dimension of vertex representation; default is 64
    -undirected <int>
        Whether the edge is undirected; default is 1
    -negative_samples <int>
        Number of negative examples; default is 5
    -window_size <int>
        Size of skip-gram window; default is 5
    -walk_times <int>
        Times of being staring vertex; default is 10
    -walk_steps <int>
        Step of random walk; default is 40
    -threads <int>
        Number of training threads; default is 1
    -alpha <float>
        Init learning rate; default is 0.025
    -use_structure <int>
        Whether using the structural information or not; default is 0

For example, if you have a user-item weighted-edges dataset named barbell.edgelist, you can use the following command to train the dataset with structural information.

./HPE -train barbell.edgelist -save embedding.embed -dimensions 64 -undirected 1 -use_structure 1