Context Vector Data Description (CVDD): An unsupervised anomaly detection method for text

This repository provides a PyTorch implementation of Context Vector Data Description (CVDD), a self-attentive, multi-context one-class classification method for unsupervised anomaly detection on text as presented in our ACL 2019 paper.

Citation and Contact

You find the ACL 2019 paper at https://www.aclweb.org/anthology/P19-1398.

If you find our work useful, please also cite the paper:

@inproceedings{ruff2019,
  title     = {Self-Attentive, Multi-Context One-Class Classification for Unsupervised Anomaly Detection on Text},
  author    = {Ruff, Lukas and Zemlyanskiy, Yury and Vandermeulen, Robert and Schnake, Thomas and Kloft, Marius},
  booktitle = {Proceedings of the 57th Conference of the Association for Computational Linguistics},
  month     = {jul},
  year      = {2019},
  pages     = {4061--4071}
}

If you would like to get in touch, just drop an email to contact@lukasruff.com.

Abstract

There exist few text-specific methods for unsupervised anomaly detection, and for those that do exist, none utilize pre-trained models for distributed vector representations of words. In this paper we introduce a new anomaly detection method---Context Vector Data Description (CVDD)---which builds upon word embedding models to learn multiple sentence representations that capture multiple semantic contexts via the self-attention mechanism. Modeling multiple contexts enables us to perform contextual anomaly detection of sentences and phrases with respect to the multiple themes and concepts present in an unlabeled text corpus. These contexts in combination with the self-attention weights make our method highly interpretable. We demonstrate the effectiveness of CVDD quantitatively as well as qualitatively on the well-known Reuters, 20 Newsgroups, and IMDB Movie Reviews datasets.

Installation

This code is written in Python 3.7 and requires the packages listed in requirements.txt.

Clone the repository to your machine and directory of choice:

git clone https://github.com/lukasruff/CVDD-PyTorch.git

To run the code, we recommend setting up a virtual environment, e.g. using virtualenv or conda:

virtualenv

# pip install virtualenv
cd <path-to-CVDD-PyTorch-directory>
virtualenv myenv
source myenv/bin/activate
pip install -r requirements.txt

conda

cd <path-to-CVDD-PyTorch-directory>
conda create --name myenv
source activate myenv
while read requirement; do conda install -n myenv --yes $requirement; done < requirements.txt

After installing the packages, run python -m spacy download en to download the spaCy en library.

Running experiments

You can run CVDD experiments using the main.py script.

The following are examples on how to run experiments on Reuters-21578, 20 Newsgroups, and IMDB Movie Reviews as reported in the paper.

Reuters-21578

Here's an example on reuters with 'ship' (--normal_class 6) considered to be the normal class using GloVe_6B word embeddings for a CVDD model with --n_attention_heads 3 and --attention_size 150.

cd <path-to-CVDD-PyTorch-directory>

# activate virtual environment
source myenv/bin/activate  # or 'source activate myenv' for conda

# change to source directory
cd src

# create folder for experimental output
mkdir ../log/test_reuters

# run experiment
python main.py reuters cvdd_Net ../log/test_reuters ../data --device cpu --seed 1 --clean_txt --embedding_size 300 --pretrained_model GloVe_6B --ad_score context_dist_mean --n_attention_heads 3 --attention_size 150 --lambda_p 1.0 --alpha_scheduler logarithmic --n_epochs 100 --lr 0.01 --lr_milestone 40  --normal_class 6;

The indexation of classes is [0, 1, 2, 3, 4, 5, 6] for ['earn', 'acq', 'crude', 'trade', 'money-fx', 'interest', 'ship'].

20 Newsgroups

Here's an example on newsgroups20 with 'comp' (--normal_class 0) considered to be the normal class using FastText_en word embeddings for a CVDD model with --n_attention_heads 3 and --attention_size 150.

cd <path-to-CVDD-PyTorch-directory>

# activate virtual environment
source myenv/bin/activate  # or 'source activate myenv' for conda

# change to source directory
cd src

# create folder for experimental output
mkdir ../log/test_newsgroups20

# run experiment
python main.py newsgroups20 cvdd_Net ../log/test_newsgroups20 ../data --device cpu --seed 1 --clean_txt --embedding_size 300 --pretrained_model FastText_en --ad_score context_dist_mean --n_attention_heads 3 --attention_size 150 --lambda_p 1.0 --alpha_scheduler logarithmic --n_epochs 100 --lr 0.01 --lr_milestone 40 --normal_class 0;

The indexation of classes is [0, 1, 2, 3, 4, 5] for ['comp', 'rec', 'sci', 'misc', 'pol', 'rel'].

IMDB Movie Reviews

Here's an example on training a CVDD model with --n_attention_heads 10 and --attention_size 150 on the full imdb training set (selected via --normal_class -1) using GloVe_42B word embeddings.

cd <path-to-CVDD-PyTorch-directory>

# activate virtual environment
source myenv/bin/activate  # or 'source activate myenv' for conda

# change to source directory
cd src

# create folder for experimental output
mkdir ../log/test_imdb

# run experiment
python main.py imdb cvdd_Net ../log/test_imdb ../data --device cpu --seed 1 --clean_txt --embedding_size 300 --pretrained_model GloVe_42B --ad_score context_dist_mean --n_attention_heads 10 --attention_size 150 --lambda_p 10.0 --alpha_scheduler soft --n_epochs 100 --lr 0.01 --lr_milestone 40 --normal_class -1;

Have a look into main.py for all the possible arguments and options.

License

MIT