Bandit-based Filtering Algorithm for Large-scale Real-time Inspection
Abstract
This project considers a large-scale real-time inspection system whose goal is to detect critical incidents out of a massive stream of events- e.g., online detection of frauds, abusers, or spams from a sequence of financial transactions, user activities, or twitter messages. When the system’s resource is not sufficient enough to inspect every single event, it is crucial to decide which events to inspect in real-time. This is a screening task as an online stochastic optimization problem, and suggest to utilize the ideas developed in the bandit literature to develop effective filtering algorithms. The bandit-based filtering algorithms will learn their own filtering rules and adaptively adjust them over time as the inspection results reveal. The algorithms will be tested using a real-world dataset for the task of detecting fraudulent ad click traffic in mobile apps, and also using a synthetic dataset in order to evaluate their scalability and robustness
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 22, 2024
- Source ID
- FA23862314122
Entities
People
- Seungki Min
Organizations
- Air Force Office of Scientific Research
- KAIST
- United States Air Force