This repository contains a Python package that provides the following code:

• 🔥 PyTorch implementation of Janus $\beta\text{-VAE}$ and Janus DDM;
• ⚡️ Janus-* training script using PyTorch Lightning;
• 👨‍💻 Hypergraph Analysis script to compare generated vs target dataset hypergraphs;
• 📄 Results report script to construct table and charts from generation results;
• ⌨️ A simple command line interface that can be used to launch training and sampling scripts.

Setup

pip install git+https://github.com/anonymous-author-1234/Janus

Requirements. Additionally to the code provided in this repository, you should install torch and lightning.

Training

janus [DATASET] <OPTIONS> train <OPTIONS> bvae <OPTIONS> --vertex-encoding/--no-vertex-encoding

and then

janus [DATASET] train <OPTIONS> ddm <OPTIONS> --bvae-ckpt

Model size

The size of both $\beta\text{-VAE}$ and DDM can be set using the --model-size option of the janus train command. The following models sizes are available.

Table 1. Model sizes available for the BETA-VAE.

Size	# Layers	# Heads	Hidden size
S	1	2	128
M	2	4	256

Table 2. Model sizes available for the DDM.

Size	# Layers	# Heads	Hidden size
S	2	4	256
M	3	8	512

Model hyperparameters

We trained both models fixing the following hyperparameters.

• Shared Hyperparameters
  • batch size 32
  • samples per hyperedge 2
  • biased random walk p 2.0
  • biased random walk q 0.5
• $\beta\text{-VAE}$
  • patience 100
• DDM
  • training timesteps 1000

The learning rate is determined using a learning rate finder strategy.

Analysis

janus-stats analysis <OPTIONS>

Compare

First of all you need to download the datasets and analyze them:

janus-stats parse [DATASET_NAME]
janus-stats analyze --root-dir references

At this point, you can perform analysis over all hypergraphs in the samples/ directory:

janus-stats analyze

Finally, run the comparison script to compare each sample with its relative reference.

janus-stats compare

Reproducibility

Training

In order to run the training configuration we used in our paper, run the following commands to train both models:

janus --batch-size 32 [DATASET] train --model-size S bvae
janus --batch-size 32 --val-size 0 [DATASET] train --model-size S --samples-per-hyperedge 2 ddm --bvae-ckpt logs/[DATASET]/BVAE-Janus-S/logs/version_0/checkpoints/best.ckpt

Sampling

janus --batch-size 32 [DATASET] sample --ckpt logs/[DATASET]/logs/version_0/DDM-Janus-S/checkpoints/last.ckpt ddm

Datasets

Dataset	# Nodes	# Hyperedges	# Connected Components	LCC Ratio
anonymous-author-1234/email-Enron	143	1,512	1	1
anonymous-author-1234/email-Eu	998	25,791	20	0.9809619238476954
anonymous-author-1234/contact-high-school	327	7,937	1	1
anonymous-author-1234/contact-primary-school	242	12,799	1	1
anonymous-author-1234/NDC-classes	1,161	1,222	183	0.540913006029285

Dataset structure

data/ full-00000-of-00001.jsonl

When running the training script full is the default split for train/val/predict (the predict split is only used when performing sampling from the $\beta\text{-VAE}$). If the split is equal for train/val, the --val-size option is used to perform train/val split across that dataset, otherwise this option is ignored.

janus anonymous-author-1234/email-Enron train --train-split train --val-split val --predict-split predict bvae

Use different datasets

Our pipeline is able to handle single and multi-hypergraph datasets retrieved from huggingface. In order to work, an hypergraph dataset must be stored in a .jsonl file, where each entry is a JSON following the HIF standard:

{
  "network-type": ...,
  "metadata": ...,
  "incidences": [...],
  "nodes": [...],
  "edges": [...]
}

In order to provide nodes and edges features, these must be set in the attrs dictionary of each node/edge.

Experiments Summary

Model/Dataset	email-Enron	email-Eu	NDC-classes	contact-high-school	contact-primary-school
BVAE	X	X	X	X	X
DDM	X	/	X	X	X
BVAE-V	X	/	X	X	X
DDM-V	-	-	-	-	-