nokiroki / out-of-distribution-ts-detection

Diploma project for my university

Out-Of-Distribution Detection for Event Sequences

Student: Belousov Nikita

Scientific advisor: Zaytsev Alexey

Problem

In this project we consider a challenging problem of anomaly detection algorithms. We will concentrate on distinguishing abnormal financial data for each unique user.

Our main approach will be to use autoencoding techniques for data reconstruction. As a result we hope to get abnormal loss growth for the anomaly samples.

Requirements

All the requirements are listed in requirements.txt

For install all packages run.

pip install -r requirements.txt

Configs

Config	Location
Main config	config.yaml
Datasets	datasets
Embedding models	embed_model
Models	model(UNDER CONSTRUCTION)

IMPORTANT! For now repo is on construction. Config launch is available only for tr2vec feature.

Data

You can find all of the necessary data in here.

Now only one dataset is available (new_data). To begin experiments you need to place transactions.parquet file into data/new_data directory.

Experiments

To launch experiments simply run the following command:

python main.py

with necessary config parameters.

Logs, model and results you can find on my commet (tags with diploma suffix) here.

Results

As the result, we trained two autoencoder models and meta-classifier for distinguishing anomaly transactions. Results we've got tell us about possibility of out method. So, as our future work with this project we will be moving rapidly toward the GAN methods.