Scalable online change-points and events detection for dynamic structured data – SCOPED

Change-point detection (CPD) is a critical problem in statistics and machine learning that focuses on identifying abrupt changes in data properties over time. These shifts, known as change-points, indicate transitions in the underlying distribution or dynamics of a system. The objective of CPD is to pinpoint when these changes occur and, in some cases, to understand the nature of the shifts. CPD has diverse applications in areas requiring real-time insights, such as medical monitoring, trading, and network security. A growing number of these applications are generating structured, high-dimensional data with non-trivial and intricate geometric properties. These data often display complex relationships and dependencies that go beyond Euclidean spaces, necessitating sophisticated techniques for analysis and interpretation. Prominent examples include time sequences on groups and manifolds, time sequences of graphs, and graph signals, all of which are central to this project.
A key challenge is finding effective representations for these data types, which often involve complex geometries not easily captured by Euclidean spaces. Another challenge is the sheer volume of data generated by sensors and infrastructures, which requires online processing to make timely decisions. The SCOPED project aims to develop unsupervised online CPD algorithms tailored for dynamic structured data, with a focus on non-parametric ML-based methods that minimize assumptions about data distributions. These algorithms will be scalable and capable of detecting a broad range of changes. Additionally, an open-world recognition setting will be developed to classify changes on-the-fly into either previously seen classes or unknown classes. Applications to smart cities monitoring and geospatial navigation are considered.