About — STAND: A Self-Tuning Anomaly Detection System

Overview

STAND is a self-tuning anomaly detection system designed to address the challenges of method selection and hyper-parameter tuning while remaining unsupervised. STAND frees users from the tedious manual tuning process often required for anomaly detection by intelligently identifying high likelihood inliers and outliers. STAND features a responsive visual interface allowing for seamless interaction, providing the user with insightful knowledge of how STAND operates. STAND outperforms the best unsupervised anomaly detection methods, yielding results similar to supervised methods that have access to ground truth labels.

Recently, we modernized and substantially expanded the STAND web platform to improve its usability, scalability, and maintainability. Specifically, we redesigned the user interface through an iterative, user-centered design process, added support for rerunning experiments with customizable detector parameters, implemented run history, downloadable logs, and robust error handling, and enabled multiple concurrent users through session-based management. On the backend, we refactored the application into a layered architecture using Flask blueprints and the factory design pattern, restructured the database to support multi-user execution, developed comprehensive automated test suites, and optimized performance by improving SQL queries and batch-processing datasets.

Project website Original algorithm implementation Original demo source code Current system repository

How it works

Getting started

1

Upload your dataset

STAND currently accepts CSV and ARFF files.

2

Choose your methods and columns

Select the unsupervised outlier detection methods to run, then provide the id and label column names from your dataset.

3

Set your expected outlier range

Input the minimum and maximum percentage of outliers you expect are contained within your dataset.

System design

STAND architecture

Sample run

Example output on the Thyroid dataset

People

Our team

Faculty & researchers

Dr. Elke Rundensteiner Professor of Computer Science, Worcester Polytechnic Institute

Dr. Samuel Madden Professor of Computer Science, Massachusetts Institute of Technology

Dr. Lei Cao Assistant Professor of Computer Science, University of Arizona

Lei Ma PhD Student in Data Science, Worcester Polytechnic Institute

Dennis Hofmann PhD Student in Data Science, Worcester Polytechnic Institute

Peter VanNostrand PhD Student in Data Science, Worcester Polytechnic Institute

Josh DeOliveira PhD Student in Data Science, Worcester Polytechnic Institute

Student contributors

Talia Andrews MS Student in Computer Science, Worcester Polytechnic Institute

Tristan Sharich BS Student in Computer Science, Worcester Polytechnic Institute

James Yi BS Student in Computer Science, Worcester Polytechnic Institute

Ishayu Das MS Student in Computer Science, Worcester Polytechnic Institute

Citations

Original work

AutoOD Demo: Hofmann, Dennis, Peter VanNostrand, Huayi Zhang, Yizhou Yan, Lei Cao, Samuel Madden, and Elke Rundensteiner. "A demonstration of AutoOD: a self-tuning anomaly detection system." Proceedings of the VLDB Endowment 15, no. 12 (2022): 3706-3709.
github.com/dhofmann34/AutoOD_Demo

AutoOD: Cao, Lei, Yizhou Yan, Yu Wang, Samuel Madden, and Elke A. Rundensteiner. "AutoOD: Automatic Outlier Detection." Proceedings of the ACM on Management of Data 1, no. 1 (2023): 1-27.
github.com/yizhouyan/AutoOD

Funding

Acknowledgments

This work was supported in part by NSF under grants IIS1910880, CSSI-2103832, CNS-1852498, NRT-HDR-1815866 and by the U.S. Dept. of Education under grant P200A180088.