Statistical Learning & Inference Seminars

The seminars will take place every Tuesday 11am-12pm during term time.

Special Pre-term seminar

Date, Time and Room

Speaker

Title

24/09, 11am, B3.02 (Zeeman)

Botond Szabo (Bocconi)

Vecchia approximation for Deep Gaussian Processes

Abstract:

Deep Gaussian processes have several advantages compared to standard Gaussian Processes (GPs), including learning local and compositional structures of the signal. However, their practical applicability is hindered by their computational complexity and instability of approximation methods. In this work we propose a novel Vecchia approximation of Deep GPs, derive optimal contraction rates for structured (compositional) functions and discuss algorithmic aspects. We also demonstrate the applicability of the proposed method on synthetic data sets. Based on an ongoing joint work with Ismael Castillo (Sorbone, Paris), Thibault Randrianarisoa (Vector Institute, Toronto) and Yichen Zhou (University of Hong Kong).

Term 1, 26-27

Date, Time and Room

Speaker

Title

06/10, 11am, MB2.22

Andrew Yiu (Southampton)

Approximating causal effects under positivity violations

Abstract:

Positivity violations constitute one of the two principal challenges of causal inference from observational data, alongside unmeasured confounding. The positivity assumption requires every unit to have a strictly positive probability of receiving each treatment, conditional on the covariates used for adjustment. When positivity is violated, there are regions of the covariate space in which no outcome information is available under one of the treatments. In the most serious case, the target causal effect is not identifiable from the observed data, leaving us either to change the target estimand or to rely on untestable extrapolation assumptions. We introduce a new method called "overlap approximation", which produces a class of intermediate estimands that progressively approach the target while remaining identifiable and efficiently estimable under positivity violations. We discuss uniform semiparametric inference for these classes, including the use of bootstrapping to construct confidence bands and tests.

13/10, 11am, MB2.22

Vaidehi Dixit (Nottingham)

Abstract:

20/10, 11am, MB2.22

Matt Nunes (Bath)

Multiscale dynamic dependence estimation over networks

Abstract:

Many multivariate time series datasets naturally exhibit nonstationary characteristics. In addition, in many applications the data are associated to a network structure, and thus analysis tools should incorporate dependence which is restricted to this structure. In this work, we propose a new modelling framework for data arising on networks which capture data nonstationarities. We introduce the concept of a *local* partial correlation graph, which formalises the lack of dependence between components of the series to non-edges in the time-varying inverse wavelet spectrum, and allows us to estimate model quantities consistently via a wavelet-spectral graphical modelling approach. We illustrate our proposed model on data arising from assets in a network of banking institutions.

27/10, 11am, MB2.22

Irène Gijbels (KU Leuven)Link opens in a new window

Abstract:

03/11, 11am, MB2.22

Yunxiao Chen (LSE)Link opens in a new window

Abstract:

10/11, 11am, MB2.22

Rowland Seymour (Birmingham)

Bayesian Computation for Spatial Comparative Judgement Models

Abstract:

Comparative judgement studies elicit quality assessments through pairwise comparisons, typically analysed using the Bradley–Terry model. We have developed a suite of spatial comparative judgement models and run studies to map the prevalence of crimes such as domestic abuse, modern slavery, and forced marriage. This talk describes work on extracting more information per comparison, organised around three extensions to the standard model. First, a multivariate normal prior distribution whose covariance matrix is built from the network of areas in the study. This lets neighbouring areas borrow strength and reduces the data required by around ninety per cent. Second, a Pólya-Gamma latent variable representation that yields a Gaussian full conditional while retaining that correlated prior distribution, which allows inference to be carried out in 20 to 30 seconds. Third, an experimental design that weights pairs by the prior variance they explain, made computable for large studies by a reduced basis decomposition of the associated pairs-of-pairs covariance matrix. We demonstrate these methods in several real world studies, describe the changes they have brought about in safeguarding practice, and outline our steps towards commercialising the work.

17/11, 11am, MB2.22

Sarah Heaps (Durham)

Abstract:

24/11, 11am, MB2.22

Sze Ming (Arthur) Lee (LSE)

Abstract:

01/12, 11am, MB2.22

Amanda Lenzi (Edinburgh)Link opens in a new window

Abstract:

08/12, 11am, MB2.22

Abstract:

Term 2, 26-27

Date, Time and Room

Speaker

Title

12/01, 11am, MB0.08

Anders Kock (Oxford)

Abstract:

19/01, 11am, MB2.22

Kartik Waghmare (ETH Zurich)

Abstract:

26/01, 11am, MB2.22

Alex Gibberd (Heriot-Watt)

Abstract:

02/02, 11am, MB2.22

Chao Zheng (Southampton)

Abstract:

09/02, 11am, MB2.22

Abstract:

16/02, 11am, MB2.22

Qingyuan Zhao (Cambridge)

Abstract:

23/02, 11am, MB2.22

Oliver Dukes (Ghent) TBC

Abstract:

02/03, 11am, MB2.22

Abstract:

09/03, 11am, MB2.22

Abstract:

16/03, 11am, MB2.22

Abstract:

Term 3, 26-27

Date,Time and Room

Speaker

Title

27/04, 11am MB0.07

Abstract:

04/05, 11am MB0.07

Abstract:

11/05, 11am MB0.07

Abstract:

18/05, 11am MB0.07

Abstract:

25/05, 11am MB0.07

Abstract:

01/06, 11am MB0.07

Abstract:

08/06, 11am MB0.07

Abstract:

15/06, 11am MB0.07

Abstract:

22/06, 11am MB0.07

Abstract:

29/06, 11am MB0.07

Abstract:

Past seminars of this series

2025/26 seminars

2024/25 seminars

2023/24 seminars