Exchangeable Gaussian Processes with application to epidemics
summary
The gist
Exchangeable Gaussian Processes with application to epidemics develops a Bayesian non-parametric framework based on multi-task Gaussian processes appropriate for temporal shrinkage to obtain…
In short
This research developed an exchangeable Gaussian Process framework using multi-task GPs to model infectious disease spread over time. The method uses hierarchical structures to learn a common underlying mean process, allowing for evidence-based knowledge of disease burden. The resulting models show improved predictive accuracy compared to simpler alternatives.
Key concepts
- Exchangeable Gaussian Processes
- This is a Bayesian non-parametric framework that uses Gaussian Processes (GPs) to model data where the relationship between different groups or tasks is exchangeable. This means the structure of dependence between groups can be captured by a shared underlying process, allowing for statistical shrinkage towards this common trend.
- Multi-task Gaussian Processes
- This approach involves using multiple GPs simultaneously, often applied across different levels of hierarchy (like countries or age groups). It allows the model to learn complex relationships and incorporate information from various data sources while maintaining a unified probabilistic structure for modeling time-dependent processes.
- Latent Stochastic Process ($lj(t)$)
- This represents an unobserved, underlying stochastic process that drives the observed epidemic data. In this model, the common mean process $lj(t)$ is postulated to be a Gaussian Process with zero mean, which serves as the central tendency around which task-specific variations are measured.
- Predictive Distribution
- This is the probabilistic forecast for future observations. It is derived using formulas involving the covariance function of the GP and historical data. This distribution provides not just a single prediction but a range of likely outcomes, helping to quantify uncertainty in future disease incidence.
Terminology used across episodes
This episode discusses
- Exchangeable Gaussian Processes with application to epidemics · Paper Radio
- On robust Bayesian causal inference
- On a Reinforcement Learning Methodology for Epidemic Control, with application to COVID-19
- Outlier-Robust Multi-Output Gaussian Processes · Paper Radio
The paper
Exchangeable Gaussian Processes with application to epidemics · Read on arXiv
Lampros Bouranis, Petros Barmpounakis, Nikolaos Demiris, Konstantinos Kalogeropoulos
Department of Statistics, Athens University of Economics and Business · Department of Oncology, University of Cambridge · Department of Statistics, The London School of Economics and Political Science
We develop a Bayesian nonparametric framework for modeling collections of related epidemic trajectories using exchangeable Gaussian processes. The proposed formulation induces structured dependence across time-varying transmission processes through shared latent Gaussian process components, leading to interpretable pooling behaviour, variance decompositions and intra-class correlation parameters that quantify information sharing across populations. The framework accommodates both direct and indirect observation regimes, including noisy measurements of effective reproduction numbers, latent transmission models for Chikungunya outbreaks and age-structured epidemic systems driven by partially observed infection dynamics. We discuss connections with broader multi-output Gaussian process formulations, including coregionalisation and latent factor models, while emphasising the interpretation of the proposed approach as a functional random-effects framework for related temporal trajectories. A simulation study investigates the behaviour of the proposed models under varying cross-population dependence structures including dependent, weakly dependent and independent settings. Across several epidemic applications, the proposed models demonstrate strong predictive performance while providing interpretable insights into shared and population-specific transmission dynamics.
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Exchangeable Gaussian Processes with application to epidemics".
Jane: Here's the gist of the paper. Exchangeable Gaussian Processes with application to epidemics develops a Bayesian non-parametric framework based on multi-task Gaussian processes appropriate for temporal shrinkage to obtain evidence-based knowledge of infectious disease burden <ref:2512.05227#pg2> The proposed models result in improved predictive ability against competing alternatives <ref:2512.05227#pg7> Model Construction The standard version of the model can be constructed in two steps, assuming an underlying latent stochastic process µ(t) and defining the stochastic process Di(t) as the cross-sectional differences between xi(t) and µ(t) <ref:2512.05227#pg5> The common mean process µ(t) is postulated to be a Gaussian Process with zero mean, although this can be easily relaxed <ref:2512.05227#pg4> The model structure allows for the incorporation of multiple hierarchical layers, which may stem from data across different countries, regions within each country, or age groups within each region/country <ref:2512.05227#pg7> Likelihood and Inference The model is parameterized as π(X,θ Y) ∝ f(Y X,θ)π(X θ)π(θ), where f(·) denotes the likelihood <ref:2512.05227#pg8> Inference on M can be carried out through π(M X,θ,Y) = π(M X,θ), which is given by a specific posterior distribution formula <ref:2512.05227#pg9> For linear and Gaussian likelihoods, the marginal likelihood f(Y θ) can be derived as Y θ ∼ N(0,K +σ 2yIn) The predictive distribution for future observation Y∗ is given by a formula involving KY∗Y and the covariance function Epidemic Dynamics The framework describes an exchangeable dynamic modelling framework appropriate for inference on the transmission process conditional on certain observed outcomes In a hierarchical transmission model with covariates, the logit of the island-specific coefficient is modeled as a Gaussian Process with structured noise The observation process for incidence counts is modeled using an over-dispersed count model where V
ys,t: = ds,t(1 + φO) A GP-driven multi-type epidemic model adapts a discrete-time age-specific renewal model by assigning multi-task GPs to logit(β1,t),…, logit(βA,t) Application and Validation The application to real data involves using Markov chain Monte Carlo (MCMC) algorithms like the Hamiltonian Monte Carlo algorithm of the Stan software for inference Prequential analysis on COVID-19 reproduction number in Europe showed that hierarchical models driven by squared exponential kernel GPs exhibit better predictive ability compared to those driven by Brownian motions For Chikungunya epidemics, the xBM model stood out with small differences to other GP-driven models in the information criteria estimates The exchangeable and multiple exchangeable models have better predictive ability compared to the country-specific model fits The xBM model was independently validated using age-stratified numbers of cumulative infections in England from the random contact prevalence survey The model supports prediction for both seen and unseen tasks such as countries, groups or individuals Discussion The proposed flexible multi-task GP-based model serves as an alternative to established approaches like the intrinsic coregionalisation model and the semi-parametric latent factor model by following the hierarchical structure often present in multitask settings The exchangeable GP formulation captures features explicitly, yielding predictive distributions that automatically incorporate the distinction between seen and unseen tasks This framework provides a principled way to learn about the temporal dependence structure across tasks by examining the posterior distribution of intra-task correlation coefficients The model allows for targeted investigation of specific countries or groups to assess their deviation from the common latent mean process The methodology extends naturally to longitudinal data analysis and multivariate time series, and makes it feasible to address causal inference tasks Acknowledgments The authors are grateful to Michalis Titsias for helpful discussions on Gaussian Process Theory and Computation The authors were supported by the European Union’s Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 101027218 The conflict of interest was declared as none declared The reader is referred to the online Supplementary Material for additional information regarding this work References
Title and authors: Jane: Let's talk about who wrote this and what the title actually means for us as listeners. The paper is titled "Exchangeable Gaussian Processes with application to epidemics," and the authors are Lampros Bouranis, Petros Barmpounakis, Nikolaos Demiris, and Konstantinos Kalogeropoulos.
Tom: I don't know much about their specific backgrounds right now, but what I do know is that they’re applying a sophisticated machine learning technique—these multi-task Gaussian processes—to infectious disease modeling. It’s about using statistical methods to understand transmission dynamics, which is huge for public health forecasting.
Lu: They are motivated by the hierarchical structures that we often see in data across different countries, or regions within a country, or even age groups within those regions. This structure allows them to capture cross-dependence between groups in a principled way.
Meng: So if I had to summarize the title for someone who doesn't read papers, it means they are using these exchangeable processes to model how diseases move around different populations at the same time. It sounds very abstract, so what's the practical payoff?
Lalam: The paper is proposing a way to capture cross-dependence between groups while retaining an interpretation based on an underlying mean process that itself is expressed as a Gaussian process. That’s basically giving us a statistical foundation for why certain trends are happening across different epidemic scenarios.
Jane: Right, so it’s not just about predicting one number; it's about understanding the relationship between those numbers when they come from different sources, like different countries or age cohorts. This sets the stage for how we can learn more robust patterns.
The paper's summary: Tom: So, looking at the summary of this paper, it’s really about developing a family of multi-task GP-based models that are appropriate for temporal shrinkage. Temporal shrinkage is just finding the underlying long-term trend or pattern when you look at time series data.
Lu: They focus on a particular class of dynamic hierarchical models that allow them to obtain evidence-based knowledge about infectious disease burden. This suggests they aren't just fitting noise; they are trying to extract meaningful information from the data structure itself.
Jane: What they do is build this model in two steps: first, assume there’s an underlying latent stochastic process for all the different processes we observe, and second, define a process called the cross-sectional differences between those observed processes and that common mean process.
Meng: That sounds complicated to set up initially. If you have to assume a common mean process for something that's supposed to be varying across groups, how do you handle that assumption?
Lalam: They postulate that this common mean process is a Gaussian Process with zero mean, but they also say it can be easily relaxed if the data demands it. This flexibility is key because real-world disease spread isn't always perfectly governed by one simple underlying trend.
Tom: And they are using this framework to analyze different types of outbreak data from recent epidemics to see how well their models perform compared to other methods. They found that these proposed models result in improved predictive ability against competing alternatives.
The paper's improvements: Jane: Now, let’s look at the specific improvements they suggest in this paper because that’s where the actual scientific gain is usually hidden. They focus on developing a family of multi-task GP-based models that are appropriate for temporal shrinkage.
Lu: The main improvement is that these processes relax the temporal independence assumption used in some prior work, and they include as special cases the models from other relevant research, like Leroy et al. and Seymour et al.
Tom: So they’re making sure their model isn't too rigid; it lets it adapt better to how time series data actually behave when you’re looking at multiple related things simultaneously. It moves beyond assuming everything happens independently over time.
Meng: What about the structure itself? They define the stochastic process Di(t) as the cross-sectional differences between the observed processes and this common mean process, which is also modeled as a GP. That sounds like they’re isolating the noise or variation from that main trend.
Lalam: They are modeling these differences, Di(t), as GPs too, which means they can analyze how those differences themselves have a temporal structure. This helps them understand the dynamics of how things change between groups over time.
Jane: So, to put it plainly for the listener, they’re taking complex data from many different sources and instead of modeling each source separately, they are using this exchangeable GP idea to learn one common underlying pattern that links them together across time.
Conclusion: Tom: So wrapping up this discussion on "Exchangeable Gaussian Processes with application to epidemics," the main implication is that we have a more principled way to learn about the temporal dependence structure across different tasks or groups by examining the posterior distribution of intra-task correlation coefficients.
Lu: They provide a flexible multi-task GP-based model that captures features explicitly, which means when you look at predictions for seen and unseen tasks—like predicting for a new country or group—the model automatically incorporates that distinction.
Meng: From an engineering standpoint, this framework allows targeted investigation of specific countries or groups to assess their deviation from that common latent mean process. That lets us pinpoint where things are behaving unusually.
Lalam: The authors suggest that by analyzing the posterior distribution of those intra-task correlation coefficients, we can learn about how the groups are related over time, which is a big step toward better forecasting.
Jane: So for someone who just listens to the show, what does this all mean? It means we can use these exchangeable GP formulations to gain evidence-based knowledge of infectious disease burden by finding that underlying common process.
Tom: Yeah, it gives us a tool that improves predictive ability against other methods when analyzing epidemic data. We’ll keep an eye on how this framework is used in real-world applications next.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates