| |
Detecting Anomalies in Large Graphs and Sensor Data
|
Speaker (s):
Bryan HOOI
PhD Candidate,
Carnegie Mellon University
|
|
Date:
Time:
Venue:
|
|
January 9, 2019, Wednesday
10:00am - 11:00am
Meeting Room 5.1, Level 5
School of Information Systems
Singapore Management University
80 Stamford Road
Singapore 178902
|
|
ABSTRACT
How can we detect fraudulent rating manipulation in online commerce, electrical component breakdowns from electrical grid sensor data, or traffic accidents from traffic sensor data? All these problems involve detecting anomalous events in graphs or graph-structured sensor data. With the increasing availability of web-scale graphs and numerous types of high-frequency industrial, weather and environmental sensors spanning major cities, there is a growing need for algorithms which can automatically monitor such data, and flag users or events which are anomalous or of interest. This allows humans to respond in real-time to anomalous events, without having to manually inspect large amounts of data. Hence, this talk focuses on two of my approaches, "Fraudar" and "Changedar": Fraudar is an adversarially robust algorithm for detecting fraud in a graph, which we use to successfully detect follower-buying schemes in a Twitter graph. Changedar is a streaming algorithm for detecting significant events, such as traffic accidents, in graph-structured sensor data. Both algorithms scale linearly, and provide approximation guarantees. I will then describe future avenues for extending graph-based anomaly detection methods.
About the Speaker
Bryan Hooi is a Ph.D. candidate jointly in the Machine Learning and Statistics departments in Carnegie Mellon University. He received his B.S. in mathematics and M.S. in computer science at Stanford University. He studies scalable algorithms for detecting anomalous events in large graphs and time series, with applications including fraud detection, and automatic monitoring of medical, industrial, weather, and environmental sensor data. He has 2 award-winning papers at top data mining conferences (KDD, ECML-PKDD), and his work has been featured in the media (e.g. National Science Foundation, WESA 90.5 FM).
|
