Research Showcase 2022

Abstract: Probabilistic methods for classifying text form a rich tradition in machine learning and natural language processing. For many important problems, however, class prediction is uninteresting because the class is known, and instead the focus shifts to estimating latent quantities related to the text, such as affect orideology. We focus on one such problem of interest, estimating the ideological positions of 55 Irish legislators in the 1991 Dáil confidence vote, a challenge brought by opposition party leaders against the then-governing Fianna Fáil party in response to corruption scandals. In this application, we clearly observe support or opposition from the known positions of party leaders, but have only information from speeches from which to estimate the relative degree of support from other legislators. To solve this scaling problem and others like it, we develop a text modeling framework that allows actors to take latent positions on a “gray” spectrum between “black” and “white” polar opposites. We are able to validate results from this model by measuring the influences exhibited by individual words, and we are able to quantify the uncertainty in the scaling estimates by using a sentence-level block bootstrap. Applying our method to the Dáil debate, we are able to scale the legislators between extreme pro-government and pro-opposition in a way that reveals nuances in their speeches not captured by their votes or party affiliations.

Take a look at Kenneth's slides (PDF).

Abstract: Additive models with interactions have been considered extensively in the literature, using estimation methods such as maximum likelihood, Tikhonov regularization or Gaussian process regression. We present an alternative empirical-Bayes approach to selecting interaction effects using the I-prior approach (WB, 2020). Using a parsimonious specification of hierarchical interaction spaces, model selection is simplified. Furthermore, we present an efficient EM algorithm for estimating the key hyperparameters, not available for competing approaches. The EM algorithm facilitates finding the global maximizer of the marginal likelihood.

Simulations for linear regressions indicate competitive performance with methods such as the lasso and Bayesian variable selection using spike and slab priors or g-priors. However, the proposed methodology is more general and can also be used with interacting nonlinear regression functions.

The regression functions live in hierarchical interaction spaces for which we consider reproducing kernel Krein spaces (RKKSs), which generalize the too restrictive reproducing kernel Hilbert spaces (RKHSs) by loosening the positive definiteness requirement. Regardless of this, the Fisher information for the regression function is positive definite, hence defining an RKHS; loosely speaking, the norm of a function in this RKHS measures the difficulty of estimating it. Then, the I-prior maximizes entropy subject to aconstant difficulty of estimating the regression function, leading to a proper Gaussian prior, ie its paths are a.s. in the required RKKS thereby obviating the need for a representer theorem.

Take a look at Wicher's slides (PDF).

Abstract: The first project is a collaborative award funded by the Wellcome Trust entitled Evaluating Policy Implementations TO Predict MEntal health (EPITOME): a Bayesian hierarchical framework for quasi-experimental designs in longitudinal settings. I am collaborating with colleagues Gianluca Baio and James Kirkbride from UCL and Marta Blangiardo from IC.

In this project the main aim is to use the interrupted time series design for causal inference to attempt to evaluate the effect of various austerity policies introduced by the Conservative/coallition goverments in the UK starting in 2010 on the mental health of minority communities in London. We also hope to be able to get a handle on the effect of the Hostile Environment Policy, introduced in 2012, which lead to the Windrush Scandal. We will be using novel Bayesian approaches to estimate synthetic and negative outcome controls amongst others. We have access to secondary referrals to mental health services in various trusts in London.

The second project is in collaboration with a former teaching fellow of the department, now working as quantitative criminologist at Leeds, Jose Pina-Sanchez. It is entitled Exploring the Nature of Ethnic Disparities in Sentencing through Causal Inference.

This is very much an applied project which aims at using DAG models, causal inference techniques and sensitivity analysis to explore how discrimination enters the judicial system. It is apparently a wide-spread belief in the UK judicial system that there is no discrimination, and we would like to show that the data we observe is in fact congruous with the presence of discrimination and that it cannot be ruled out. We will have access to individual level sentencing data--via secure access facilities--which the UK government are making available for the first time.

Take a look at Sara's slides (PDF).

Abstract: Regression with Gaussian Process (GP) prior is a powerful statistical tool for modelling a wide variety of data with both Gaussian and non-Gaussian likelihood. In the spatial statistics community, GP regression, also known as Kriging, has a long-standing history. It has been proven useful since its introduction, due to its capability of modelling autocorrelation of spatial and spatio-temporal data.

Other than space and time, real-life applications often contain additional information with different characteristics. In applied research, interests often lie in exploring whether there exists a space-time interaction or investigating relationships with covariates and the outcome while controlling for space and time effect.

Additive GP regression allows to model such flexible relationships by exploiting the structure of the GP covariance function (kernel) by adding and multiplying different kernels for different types of covariates.This has only partially be adapted in spatial and spatio-temporal analysis.

In this study, we use ANOVA decomposition of kernels and introduce a unified approach to model spatio-temporal data, using the full flexibility of additive GP models. Not only does this permit modelling of main effects and interactions of space and time, but furthermore to include covariates, and let the effects of the covariates vary with time and space. We consider various types of outcomes including, continuous, categorical and counts. By exploiting kernels for graphs and networks, we show that areal data canbe modelled in the same manner as the data that are geo-coded using coordinates.

Take a look at Sahoko's slides (PDF).

Abstract: Tensor time series data appears naturally in a lot of fields, including finance and economics. As a major dimension reduction tool, similar to its factor model counterpart, the idiosyncractic components of a tensor time series factor model can exhibit serial correlations, especially in financial and economic applications. This rules out a lot of state-of-the-art methods that assume white idiosyncractic components, or even independent/Gaussian data.

While the traditional higher order orthogonal iteration (HOOI) is proved to be convergent to a set of factor loading matrices, the closeness of them to the true underlying factor loading matrices are ingeneral not established, or only under some strict circumstances like having i.i.d. Gaussian noises (Zhang and Xia, 2018). Under the presence of serial and cross-correlations in the idiosyncractic components and time series variables with only bounded fourth order moments, we propose a pre-averaging method that accumulates information from tensor fibres for better estimating all the factor loading spaces. The estimated directions corresponding to the strongest factors are then used for projecting the data for a potentially improved re-estimation of the factor loading spaces themselves, with theoretical guarantees and rate of convergence spelt out. We also propose a new rank estimation method which utilizes correlation information from the projected data, in the same spirit as Fan et al (2020) for factor models with independent data. Extensive simulation results reveal competitive performance of our rank and factor loading estimators relative to other state-of-the-art or traditional alternatives. A set of matrix-valued portfolio return data is also analysed.

Take a look at Clifford's slides (PDF).

Abstract: In the talk, I will discuss composite likelihood estimation and testing for latent variable models for categorical data. Latent variable models are widely used in social sciences for measuring unobserved constructs such as ability, attitudes, health state, etc. The constructs are measured using a number of observed variables (items) that can be continuous, categorical or of mixed type. The computational complexity of the models depends on the number of observed variables and latent variables. Frequentist and Bayesian estimation methods have been proposed in the literature for estimating the model parameters. In addition, full and limited information methods have been developed within the frequentist approach. One of the limited information methods is known as composite likelihood and in particular in our case pairwise likelihood estimation. I will review the pairwise likelihood framework for the models for ordinal and binary data and discuss some recent advances.

Take a look at Irini's slides (PDF).

Abstract: After presenting myself, I will tell you about the research I have done during the PhD. In the first half of the talk, you can expect stories about models for sparse networks, underpinned by Bayesian nonparametric theory, and a discussion about their asymptotic properties. In the second half I will concentrate on statistical methods to assess disclosure risk, a method to quantify the level of privacy of a non-perturbed dataset. I will wrap up briefly mentioning some work on fair generalised linear models using a simple ridge regression.

Take a look at Francesca's slides (PDF).

Abstract: For many problems in statistics, finance, and insurance, there is a significant interest in understanding so-called invariance principles, which capture the approximate behavior of a given model under suitable scaling limits. Often, this leads to questions of functional convergence for stochastic integrals. In this talk, I will address how to achieve such convergence when the models of interest have certain autoregressive features that take us outside the realm of the classical literature.

Take a look at Andreas's slides (PDF).

Abstract: Style transfer is a central problem in data science with numerous successful applications. I will present a novel style transfer framework relying on vector-valued reproducing kernel Hilbert spaces. The idea is instantiated in emotion transfer where the goal is to transform facial images to different target emotions, with explicit control over the continuous style space. The proposed approach achieves low reconstruction cost and high emotion classification accuracy on various popular facial emotion benchmarks.

Take a look at Zoltan's slides (PDF).

Abstract: Recently, reinforcement learning (RL) has attracted substantial research interests. Much of the attention and success, however, has been for the discrete-time setting. Continuous-time RL, despite its natural analytical connection to stochastic controls, has been largely unexplored and with limited progress. In particular, characterising sample efficiency for continuous-time RL algorithms remains a challenging and open problem. In this talk, we will discuss some recent advances in the regret analysis for the episodic linear-convex RL problem, and report a sublinear (or even logarithmic) regret bound for a learning algorithm inspired by filtering theory. The approach is probabilistic, involving quantifying the precise performance gap between applying greedy policies derived from estimated and true models, and exploring the concentration properties of sub-Weibull random variables.

Take a look at Yufei's slides (PDF).