Change Point Detection for Random Objects with Possibly Periodic Behavior

2025-01-03

AI Summary

This paper addresses the precise localization of distributional change points in time series of stochastic objects—such as graph Laplacian matrices—that evolve on non-Euclidean domains and exhibit periodic behavior. To overcome the limitation of existing methods, which struggle to distinguish genuine changes from spurious fluctuations induced by periodicity, we propose the first nonparametric scan statistic specifically designed for periodic stochastic objects. Leveraging an endpoint-truncation strategy and asymptotic distribution theory, we rigorously establish its detection consistency and near-optimal localization accuracy. Our method integrates graph Laplacian-based representation learning with change-point consistency analysis. Extensive experiments on synthetic data and real-world NYC Citi Bike network traffic demonstrate significant improvements over baselines: it accurately identifies interpretable change points that align closely with major historical events.

Technology Category

• Planning, Routing, and Scheduling: Optimization of Spatio-temporal Systems
• Search and Optimization: Distributed Search
• Reasoning under Uncertainty: Stochastic Optimization

Application Category

• Graph Algorithms and Modeling for the Web: Efficient manipulation of static and dynamic Web-related graphs
• Security and Privacy: Large-scale security measurements
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods

Abstract

Time-varying random objects have been increasingly encountered in modern data analysis. Moreover, in a substantial number of these applications, periodic behavior of the random objects has been observed. We introduce a new, powerful scan statistic and corresponding test for the precise identification and localization of abrupt changes in the distribution of non-Euclidean random objects with possibly periodic behavior. Our approach is nonparametric and effectively captures the entire distribution of these random objects. Remarkably, it operates with minimal tuning parameters, requiring only the specification of cut-off intervals near endpoints, where change points are assumed not to occur. Our theoretical contributions include deriving the asymptotic distribution of the test statistic under the null hypothesis of no change points, establishing the consistency of the test in the presence of change points under contiguous alternatives and providing rigorous guarantees on the near-optimal consistency in estimating the number and locations of change points, whether dealing with a single change point or multiple ones. We demonstrate that the most competitive method currently in the literature for change point detection in random objects is degraded by periodic behavior, as periodicity leads to blurring of the changes that this procedure aims to discover. Through comprehensive simulation studies, we demonstrate the superior power and accuracy of our approach in both detecting change points and pinpointing their locations, across scenarios involving both periodic and nonperiodic random objects. Our main application is to weighted networks, represented through graph Laplacians. The proposed method delivers highly interpretable results, as evidenced by the identification of meaningful change points in the New York City Citi Bike sharing system that align with significant historical events.

Problem

• Detecting abrupt distribution changes in non-Euclidean random objects
• Accommodating periodic behavior in change point detection
• Handling single or multiple change points in time-varying data

Innovation

• Nonparametric detector for non-Euclidean periodic objects
• Flexible framework accommodating various change point detectors
• Theoretical guarantees for detection power and localization accuracy