Reading Group

Speaker: Yawen Ma, Lancaster University

Title: Item pool quality control in educational testing: change point model, compound risk, and sequential detection.​

Abstract: ​​We will explore a sequential Bayesian change point detection in multiple parallel data streams, where each stream has its own change point. Once a change is detected in a stream, it is permanently deactivated. It introduces a compound risk function that captures the trade-off between maximize the normal operation of the pre-change streams, while controlling the proportion of post-change streams among the active streams at all time points. The motivating application comes from item monitoring in educational testing, but the framework is more generally applicable to parallel sequential detection problems.

Paper: Chen et al. (2021)

Speaker: Edwin Tang, University of Warwick

Title: E-detectors: a nonparametric framework for sequential change detection.​

Abstract: ​​E-detectors improve upon classical sequential change detection methods as we can control average run length when testing nonparametrically specified composite hypotheses in an online setting, without an iid assumption. We will study how to form e-detectors and its statistical properties. Its efficacy is presented in an application to tracking the performance of a basketball team over multiple seasons, where we have non-iid bounded random variables.

Paper: Shin et al., (2023)

Speaker: Yuqi Zhang, University of Bristol

Title: Sequential Nonparametric Tests for a Change in Distribution: An Application​

Abstract: ​​We will study a sequential nonparametric test for detecting changes in distribution based on windowed Kolmogorov–Smirnov statistics. The method is computationally efficient, requires no parametric assumptions, and allows rigorous analysis of false-alarm rates and detection power. We will discuss its application to radiological anomaly detection using background gamma-radiation data, highlighting improvements in time-to-detection over existing sequential methods.

Paper: Padilla et al. (2019) 

Speaker: Xianghe Zhu, London School of Economics

Title: Anomaly Detection for Dynamic Network​

Abstract: ​​We focus on anomaly detection in dynamic networks. We start with a brief survey snapshot to introduce common anomaly types and major methodological families, and then move to a theory-backed paper by Chen et al. The paper proposes an algorithm that uses a test statistic derived from Multiple Adjacency Spectral Embedding (MASE) to perform graph-level and vertex-level anomaly detection in time series of graphs. Beyond the algorithm itself, we will focus on the paper’s inferential perspective: how the test statistic is motivated under a latent space model, and what theoretical guarantees (e.g., level control/consistency and detectability) say about when anomalies can be reliably identified. The method is supported by both theoretical results and simulation evidence.

Paper: Chen et al., (2025), suvey paper:  Ranshous et al., (2015)

Speaker: Laura Baracaldo Lancheros, Lancaster University

Title: Bayesian Nonparametric Approaches to Abnormality​​

Abstract: ​​We will explore a Bayesian nonparametric framework for spatio-temporal anomaly detection. The paper combines a Hierarchical Dirichlet Process Hidden Markov Model (HDP-HMM) to infer an unknown number of temporal regimes with a Bayesian nonparametric factor analysis model that uncovers latent spatial activity patterns. Together, these components provide a flexible approach for detecting abnormal events in high-dimensional dynamic data without fixing the model complexity in advance.

​Speaker: Gengyu Xue, University of Warwick​

Title: Anomaly Detection for Multivariate Time Series through Stochastic Recurrent Neural Network​

Abstract:​ We will go through the paper “Robust Anomaly Detection for Multivariate Time Series through Stochastic Recurrent Neural Network” (link). We will start with a brief introduction to the key concepts used in the paper: Gated Recurrent Unit (GRU), Variational Autoencoder (VAE) and Planar Normalising Flows. We will then explore the network architecture of OmniAnomaly and discuss how the method effectively detects anomalies as well as anomaly interpretation. The session will conclude with an open discussion and feedback from the audience.

​Speaker: Zetai Cen, University of Bristol​

Title: Outlier detection in multivariate time series

Abstract:​ We provide a brief overview on outliers in multivariate time series, as a generalization from univariate to low-dimensional, and to briefly appreciate such a problem in high-dimensional context. Built on vector autoregressive and moving-average models, we show how certain model structures can be exploited to facilitate outlier detection. By no means our discussion will be complete, yet it should provide us more insight on outliers in complex data streams, and even so if anomaly is the object of interest.

​Speaker: Han Yan, London School of Economics and Political Science​

Title: Introduction to Transformer and Its Applications in Time Series Anomaly Detection

Abstract: This presentation provides an overview of the Transformer architecture, highlighting its core self-attention mechanism. We will explore how it enables the model to effectively capture long-range dependencies in sequential data. The discussion will then shift to its application in time-series and anomaly detection. We will examine how the Anomaly Transformer conducts unsupervised anomaly detection based on Anomaly-Attention.

Speaker: Kes Ward, Lancaster University​

Title: iForest, a point anomaly detection method useful for streaming data​​​

Abstract: We'll be covering iForest, a point anomaly detection method useful for streaming data. The original iForest paper (2008) can be found here. Physical copies may be available at the reading group. There will be a short presentation by Kes covering the necessary background, followed by an interactive discussion in small groups, followed by feeding back to the main group.

Key concepts to discuss: What methodological tricks make iForest work? How has iForest been extended and its weaknesses addressed/adapted to other settings? How has iForest gone from a research paper to being successfully and impactfully implemented in practice? What can we learn from all this that's useful for our own research?