Disruption

This repository contains the code for the embedding disruptiveness measure (EDM). You can find an extended abstract for IC2S2 here: https://munjungkim.github.io/files/ic2s2_2023_03_02_disruptiveness.pdf

pip package

We developed a pip package for calculating the embedding disurptiveness measure. You can install it by

pip install embedding-disruptiveness

The source code of the package is in here, and the example code for calculating embedding disruptiveness measure with the package is in notebooks/embedding-disruptiveness package.ipynb.

Environment setup

uv

If you are using uv, run the following to install necessary packages:

uv venv
source .venv/bin/activate
uv pip install -r requirements.txt

Snakemake

Since the current python code requires many parameters, it would be better to use snakemake tool. Once a container has been created, you can execute the following commands in the workflow directory:

snakemake '{working_dirctory}/data/original/150_5_q_1_ep_5_bs_4096_embedding/distance.npy' -j

This command will locate the rule for generating the {working_directory}/data/original/150_5_q_1_ep_5_bs_4096_embedding/distance.npy file, which corresponds to the calculating_distance rule in our Snakefile.

This rule takes the {working_directory}/data/original/150_5_q_1_ep_5_bs_4096_embedding/in.npy and {working_directory}/data/original/150_5_q_1_ep_5_bs_4096_embedding/out.npy files as inputs, and Snakemake will execute the rule responsible for generating these files, named as the embedding_all_network rule in our Snakefile.

The embedding_all_network rule executes the following command: python3 scripts/Embedding.py {input} {params.Dsize} {params.Window} {params.Device1} {params.Device2} {params.Name} {params.q} {params.epoch} {params.batch}. The parameters for this command are defined within the embedding_all_network rule as follows:

 params:
        Name = "{network_name}",
        Dsize = "{dim}",
        Window = "{win}",
        Device1 = "6",
        Device2 = "7",
        q = "{q}",
        epoch = "{ep}",
        batch = "{bs}"

Without Snakemake

Without snakemake, you can follow the following steps.

Embedding Calculation

To calculate the embedding vectors, you can use the following command line:

python3 scripts/Embedding.py {path/to/citation_network_file} {embedding_dimension} {window_size} {device1} {device2} {citation_network_name} {q_value} {epoch_size} {batch_size}

For example, you can run the command as shown below:

python3 scripts/Embedding.py /data/original/citation_net.npz 200 5 6 7 original 1 5 1024

Embedding.py will train the node2vec model and save the result of in-vectors under the path {path/to/citation_network_file}/{DIM}_{WIN}_q_{Q}_ep_{EPOCH}_bs_{BATCH}_embedding/. For instance, the above command will save in and out vectors in the path /data/original/200_5_q_1_ep_5_bs_1024_embedding/in.npy and /data/original/200_5_q_1_ep_5_bs_1024_embedding/out.npy.

Distance Calculation

Based on the embedding vectors you calculate from the above process, you can execute the following command to calculate the distance.

python3 scripts/Distance_Disruption.py distance {path/to/invectors}  {path/to/outvectors} {path/to/citation_network_file} {device name}

For example, you can run the command as shown below:

python3 scripts/Distance_Disruption.py distance /data/original/  {path/to/outvectors} {path/to/citation_network_file} {device name}