Unsupervised Embedding of Hierarchical Structure in Euclidean Space

This repository is the official implementation of Unsupervised Embedding of Hierarchical Structure in Euclidean Space.

Requirements

To install requirements:

pip install -r requirements.txt

Required Datasets will be downloaded automatically while running the code

Training

To train and evaluate the model on synthetic data, run this command:

cd synthetic/
python synthetic_experiments.py --n_class 8 --margin 8 --variance 1 --dim 100 --hidden_dim 3 --linkage_method ward --learning_rate 1e-3

The above command will give you the result for the BTGM in Figure 1 in our paper.

Pre-trained Models

We provide the pre-trained parameters for MNIST and CIFAR-25 (both in Pytorch) and reuters (from the original implementation of VaDE). You can download pretrained models here:

• Download parameters for CIFAR25 experiments to CIFAR25/parameters/ using this link
  
• Download reuters10k data to reuters/dataset/reuters10k using this link
  

Evaluation

To evaluate on MNIST, run:

cd MNIST/
python MNIST_experiments.py --linkage_method ward --embedding_method VaDE --rescaling_transform

📋The same procedure applies to CIFAR-25 dataset.

Results

Our proposed method achieves the following performance in terms of Dendrogram Purity and Moseley-Wang's objective :

Model name	MNIST	CIFAR-25	reuters	20newsgroups
VaDE+Ward+Trans.(DP)	0.886	0.128	0.672	0.251
VaDE+Ward+Trans.(MW)	0.955	0.472	0.768	0.606