Article
Statistics & Probability
Nicolas Verzelen, Magalie Fromont, Matthieu Lerasle, Patricia Reynaud-Bouret
Summary: Given a time series Y in Rn with a piecewise constant mean and independent components, this work investigates the problems of change-point detection and change-point localization. Optimal rates for both problems are characterized, and a phase transition phenomenon from a global testing problem to a local estimation problem is uncovered. The energy threshold for detection is established, and it is shown that the localization error becomes purely parametric for specific change-points. The paper also introduces two procedures that achieve these optimal rates.
ANNALS OF STATISTICS
(2023)
Article
Engineering, Industrial
Shiwang Hou, Keming Yu
Summary: This study introduces a new non-parametric cumulative sum (CUSUM) control chart method to monitor arbitrary distribution changes and diagnose detailed change types simultaneously, which outperforms other methods in detecting small changes.
INTERNATIONAL JOURNAL OF PRODUCTION RESEARCH
(2021)
Article
Statistics & Probability
Haoyun Wang, Liyan Xie, Yao Xie, Alex Cuozzo, Simon Mak
Summary: We present a new CUSUM procedure for sequential change-point detection in Hawkes networks using discrete events data. Simulation studies and an application in neuroengineering demonstrate the improved efficiency and performance of this method compared to existing methods.
Article
Statistics & Probability
Mengjia Yu, Xiaohui Chen
Summary: In this paper, a Gaussian multiplier bootstrap is introduced to calibrate critical values of CUSUM test statistics for detecting change points in independent samples generated from the mean-shift model. The proposed bootstrap CUSUM test is fully data dependent and has strong theoretical guarantees under various dependence structures and moment conditions. Estimators are derived for estimating change point locations by maximizing the l infinity-norm of generalised CUSUM statistics, with rates of convergence impacted by dimension only through logarithmic factors, allowing for consistency of estimators when dimension is much larger than sample size. A principled bootstrap-assisted binary segmentation (BABS) algorithm is proposed for multiple change point detection, with rate of convergence derived under suitable signal separation and strength conditions. Extensive simulation studies show empirical evidence in agreement with theoretical results.
JOURNAL OF THE ROYAL STATISTICAL SOCIETY SERIES B-STATISTICAL METHODOLOGY
(2021)
Article
Statistics & Probability
Ruiyu Xu, Jianguo Wu, Xiaowei Yue, Yongxiang Li
Summary: This article proposes an online structural change-point detection method for high-dimensional streaming data through dynamic sparse subspace learning and an efficient Pruned Exact Linear Time algorithm. The effectiveness of the method is demonstrated through simulation studies and a real case study.
Article
Automation & Control Systems
Gaetano Romano, Idris A. Eckley, Paul Fearnhead, Guillem Rigaill
Summary: This article introduces an online algorithm called FOCuS, which is capable of processing high-frequency observations and limited computational resources for changepoint detection. The algorithm achieves this by simultaneously running previous methods with different window sizes or different values for the size of change. The theoretical results demonstrate that the computational cost per iteration of FOCuS is logarithmic in the number of observations.
JOURNAL OF MACHINE LEARNING RESEARCH
(2023)
Article
Computer Science, Information Systems
Qunzhi Xu, Yajun Mei, George V. Moustakides
Summary: In this paper, a study on multi-stream sequential change-point detection is presented, as well as a strategy to solve this problem under sampling control constraints. The results show that a simple myopic-sampling-based sequential change-point detection strategy is second-order asymptotically optimal when the number of processes M is fixed.
IEEE TRANSACTIONS ON INFORMATION THEORY
(2021)
Article
Energy & Fuels
Sakitha Ariyarathne, Harsha Gangammanavar, Raanju R. R. Sundararajan
Summary: This paper presents a wind speed simulation method by detecting change points in multivariate nonstationary wind speed time series data. The method identifies changes in the covariance structure and decomposes the nonstationary time series into stationary segments, allowing for modeling and simulation within each segment. The proposed approach retains the statistical properties of the original time series and can be employed for analysis in power systems planning and operations. Experimental results demonstrate the capabilities of the change point detection method and its impact on wind generation in the economic dispatch problem.
Article
Computer Science, Software Engineering
Hongyan Xu, Ayten Yigiter, Jie Chen
Summary: In this paper, a new R package called onlineBcp is introduced, which is based on an online Bayesian change point detection algorithm. This package provides convenient output of probability, location, and statistics for multiple change points, as well as functions for missing value pre-treatment and normality assumption checking. Furthermore, it allows fast detection of new change points in a data stream.
Article
Economics
Cathy Yi-Hsuan Chen, Yarema Okhrin, Tengyao Wang
Summary: This article proposes a method and algorithm for monitoring changes in dynamic networks and addresses the challenges in network analysis. The effectiveness of the method is supported by numerical studies and an application to social media messages network.
JOURNAL OF BUSINESS & ECONOMIC STATISTICS
(2022)
Article
Computer Science, Artificial Intelligence
Andre Ferrari, Cedric Richard, Anthony Bourrier, Ikram Bouchikhi
Summary: This paper investigates a kernel-based change-point detection method that estimates the density ratio on consecutive time intervals. The algorithm is improved and its behavior is analyzed in terms of mean and mean square. The effectiveness of the algorithm is validated through Monte Carlo simulations and experiments on real-world data.
PATTERN RECOGNITION
(2023)
Article
Engineering, Industrial
Zhen Chen, Yaping Li, Di Zhou, Tangbin Xia, Ershun Pan
Summary: This paper proposes a two-phase Gaussian process degradation model with a change-point for products exhibiting two-phase patterns, allowing for parameter estimation and change-point detection, and deriving closed-form distributions for the first passage time and the remaining useful life, effectively capturing the characteristics and trends in degradation paths.
RELIABILITY ENGINEERING & SYSTEM SAFETY
(2021)
Article
Computer Science, Artificial Intelligence
Jianfeng Ren, Xudong Jiang
Summary: A three-step classification framework is proposed for UAV detection, which includes handling outliers and complex sample distribution, addressing unreliable feature dimensions of multi-Gaussian model, and fusing distances of samples to different clusters at different dimensionalities. This approach significantly outperforms the state-of-the-art methods in a large benchmark dataset.
PATTERN RECOGNITION
(2021)
Article
Statistics & Probability
Karim Atashgar, Naser Rafiee, Mahdi Karbasian
Summary: This paper introduces a new hybrid method for detecting change points in panel data. The method has a high sensitivity and is capable of identifying multiple change points.
COMMUNICATIONS IN STATISTICS-THEORY AND METHODS
(2022)
Article
Statistics & Probability
Michael Baron, Sergey V. V. Malov
Summary: Change-point detection methods are proposed for temporary failures or transient changes, where unexpected disorder is followed by re-adjustment and return to the initial state. Sequential and retrospective tools based on likelihood are used for detection and estimation of change-points. The accuracy of the estimated change-points is evaluated. The proposed methods provide simultaneous control of the familywise false alarm and false re-adjustment rates at pre-chosen levels.
JOURNAL OF APPLIED STATISTICS
(2023)
Article
Engineering, Electrical & Electronic
Yue Wu, Shaodan Ma, Yuantao Gu
IEEE TRANSACTIONS ON COMMUNICATIONS
(2020)
Article
Engineering, Electrical & Electronic
Chengzhu Yang, Yuantao Gu, Badong Chen, Hongbing Ma, Hing Cheung So
Summary: This passage discusses the importance of modeling many real-world problems as sparse matrix recovery from two-dimensional measurements, and introduces a neural network named 2D-LPGA to quickly reconstruct the target matrix. The theoretical analysis shows that the network can achieve linear convergence rate under certain conditions, and numerical experiments demonstrate its superiority over classical schemes.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS II-EXPRESS BRIEFS
(2021)
Article
Computer Science, Information Systems
Yuantao Gu, Yuchen Jiao, Xingyu Xu, Quan Yu
Summary: This work proposes an energy-efficient decentralized CPD algorithm with a communication strategy based on request-response and censoring scheme. Numerical simulations and experimental results demonstrate its high energy efficiency and small detection delay in IoT applications.
IEEE INTERNET OF THINGS JOURNAL
(2021)
Article
Automation & Control Systems
Lei Xing, Badong Chen, Shaoyi Du, Yuantao Gu, Nanning Zheng
Summary: Multiview subspace clustering aims to cluster data points with information from multiple sources or features, and has a wide range of applications. A novel correntropy-based multiview subspace clustering (CMVSC) method is proposed, which combines Frobenius norm and correntropy-induced metric (CIM) to optimize representation matrix structure and utilize information from multiple views.
IEEE TRANSACTIONS ON CYBERNETICS
(2021)
Article
Automation & Control Systems
Badong Chen, Lujuan Dang, Yuantao Gu, Nanning Zheng, Jose C. Principe
Summary: The article introduces a new Kalman-type filter called Minimum Error Entropy KF (MEE-KF), which uses the minimum error entropy criterion instead of MMSE or MCC. Similar to MCC-based KFs, the proposed filter is an online algorithm with a recursive process. Additionally, an MEE extended KF (MEE-EKF) is developed for performance improvement in nonlinear situations.
IEEE TRANSACTIONS ON SYSTEMS MAN CYBERNETICS-SYSTEMS
(2021)
Article
Engineering, Civil
Qi Zhang, Jiang Zhu, Yuantao Gu, Zhiwei Xu
Summary: This article explores direction of arrival (DOA) estimation in environments with heteroscedastic noise, proposing a multisnapshot variational line spectral estimation method (MVHN) that automatically estimates noise variance, nuisance parameters, and number of sources, providing uncertain degrees of DOA estimates. Variants of MVHN, MVHN-S and MVHN-A, are developed for scenarios where noise variance varies only across snapshots or antennas. Numerical experiments are conducted to demonstrate the performance of the proposed algorithms, including using a real dataset in a DOA application.
IEEE JOURNAL OF OCEANIC ENGINEERING
(2021)
Article
Engineering, Electrical & Electronic
Gen Li, Yuantao Gu, Jie Ding
Summary: A crucial problem in neural networks is selecting an architecture that balances the tradeoff between underfitting and overfitting. This study demonstrates that l(1) regularizations for two-layer neural networks can control generalization error and sparsify input dimensions. By applying an appropriate l(1) regularization on the output layer, the network can produce a tight statistical risk. Additionally, using l(1) regularization on the input layer results in a risk constraint that is not dependent on the input data dimension. The findings also suggest that training a wide neural network with suitable regularization offers an alternative bias-variance tradeoff over selecting from a candidate set of neural networks.
IEEE SIGNAL PROCESSING LETTERS
(2022)
Article
Computer Science, Information Systems
Gen Li, Yuting Wei, Yuejie Chi, Yuantao Gu, Yuxin Chen
Summary: This paper investigates the sample complexity of asynchronous Q-learning and proposes an improvement method to optimize the algorithm's performance.
IEEE TRANSACTIONS ON INFORMATION THEORY
(2022)
Article
Engineering, Electrical & Electronic
Yuchen Jiao, Xingyu Xu, Yuantao Gu
Summary: This article introduces a communication-efficient change-point detection algorithm that uses low precision quantized data. By combining differential quantization technique and error feedback technique, the communication cost can be effectively reduced without deteriorating the detection performance.
IEEE TRANSACTIONS ON SIGNAL PROCESSING
(2022)
Article
Engineering, Electrical & Electronic
Yuchen Jiao, Yuantao Gu
Summary: This paper introduces the application of machine learning techniques in communication systems and proposes a new algorithm called Decentralized Subspace Estimation (DSE) to solve the problem of subspace estimation in decentralized settings. The communication complexity of existing algorithms is improved from O(log(2)(1/epsilon)) to O(log(1/epsilon)) using the gradient tracking technique. The influence of network connectivity and data eigen-gap on communication complexity is theoretically analyzed and the effectiveness of the algorithm and the correctness of the theoretical result are verified through experiments.
IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING
(2022)
Article
Engineering, Electrical & Electronic
Zhouhao Yang, Xingyu Xu, Yuantao Gu
Summary: This paper presents a differentially private algorithm that accurately estimates the mean of a population with a given cumulative distribution function. The algorithm surpasses previous ones by being able to handle more general probability distributions and achieving greater accuracy with fewer samples for light-tailed distributions.
IEEE SIGNAL PROCESSING LETTERS
(2022)
Proceedings Paper
Computer Science, Artificial Intelligence
Gen Li, Changxiao Cai, Yuxin Chen, Yuantao Gu, Yuting Wei, Yuejie Chi
Summary: Q-learning, a key concept in reinforcement learning, has recently made progress in understanding sample efficiency, especially in the synchronous setting. This work sharpens the sample complexity of synchronous Q-learning, achieving higher efficiency and similar results for finite-horizon MDPs. The analysis reveals the effectiveness of vanilla Q-learning without requiring additional resources, providing insights into studying other variants of Q-learning.
INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 139
(2021)
Proceedings Paper
Computer Science, Artificial Intelligence
Gen Li, Yuantao Gu
Summary: The spectral method is commonly used for subspace clustering, where a random geometry graph is first constructed, followed by applying spectral clustering to obtain clustering results. This paper establishes a theory to show the power of spectral clustering and proves the efficiency of subspace clustering in fairly broad conditions. The insights developed in this paper may have implications for other random graph problems.
INTERNATIONAL CONFERENCE ON MACHINE LEARNING, VOL 139
(2021)
Proceedings Paper
Computer Science, Information Systems
Kunzan Liu, Yuchen Jiao, Ye Jin, Xu Xiang, Yuantao Gu
Summary: This paper proposes a new outlier detection algorithm DrSOD, which uses random projection to avoid direct access to raw data, improving data security and privacy protection capability. Theoretically proven to correctly detect outliers with overwhelming probability under connectivity assumptions, the random projection step also enhances the computational efficiency of the algorithm. Experiments on synthetic and real-world datasets demonstrate the effectiveness and efficiency of DrSOD.
2021 IEEE 5TH INTERNATIONAL CONFERENCE ON CRYPTOGRAPHY, SECURITY AND PRIVACY (ICCSP)
(2021)
Proceedings Paper
Engineering, Electrical & Electronic
Yuchen Jiao, Yirong Ma, Yuantao Gu
2020 IEEE 11TH SENSOR ARRAY AND MULTICHANNEL SIGNAL PROCESSING WORKSHOP (SAM)
(2020)