Research Interests

Machine learning (ML) has seen rapid advancements in recent years, significantly enhancing its generative and predictive capabilities. These models clearly demonstrate computational intelligence—the ability to solve complex computational and mathematical problems by leveraging data-driven learning, pattern recognition, and optimization techniques. However, the lack of transparency and the overconfidence in these predictions have raised serious concerns about their reliability and safety in real-world applications. Moreover, the challenge of designing models that align with human and societal values remains unresolved.

These issues can be attributed to the lack of epistemic intelligence in current ML systems. Epistemic intelligence refers to the ability to appreciate the limits of one’s knowledge, be honest about what is known, and navigate uncertainty and ambiguity regarding what is unknown. Effective communication and integration of knowledge is also a key component of epistemic intelligence.

My research focuses on advancing the theory and practice of epistemic machine learning by focusing on three key directions: Uncertainty-aware ML, Explainable ML, and Preference Modelling.

Uncertainty-aware ML

Learning is an inductive process where we observe information from the world, interpret and internalise it, and generalise our experiences to new scenarios. Uncertainty is an inherent part of this process, emerging at various stages: from hypothesis formulation and data collection to model optimisation, performance evaluation, and deployment. My research focuses on developing statistical methods to acknowledge, propagate, and utilise uncertainty throughout the machine learning development pipeline, thereby enabling models to make reliable and safe decisions in real-world applications.

Uncertainty in data

When data are collected at varying granularities (e.g., different image resolutions or opinion pooling versus census data), performing learning tasks on these data jointly becomes challenging. We studied this problem in our NeurIPS 2021 work "Deconditional Downscaling with Gaussian Processes", where we model coarse observations as conditional expectations of fine-grained signals, and propose a novel **Deconditional Gaussian Process**, which allows us to "undo" this aggregation procedure and subsequently perform inference. This method integrates conditional mean embeddings, a flexible kernel method that operates on distributions with minimal parametric assumptions, and Gaussian processes, a flexible model class known for effectively capturing uncertainty. We demonstrate its proficiency in an atmospheric field downscaling problem, where we show substantial improvements over existing methods.

In our NeruIPS 2021 work Uncertainty Quantfication for Causal Data Fusion", we studied the causal data fusion problem, where datasets pertaining to multiple causal graphs are combined to estiamte the average treatment effect of a target variable. As data arises from multiple sources and can vary in quality and quantity, principled uncertainty quantification becomes essential. We formulate this problem as an instance of learning over distributions, and propose a novel **Bayesian Conditional Mean Embeddings** to represent each distribution in the reproducing kernel Hilbert space (a flexible function space) while taking into account within each causal graph. The efficacy of our uncertainty is demonstrated through the task of Causal Bayesian optimisation, and show improvements over standard sampling-based methods.

Uncertainty in optimisation

Bayesian optimisation (BO) is an efficient blackbox function optimiser when function evaluations are expensive to compute and gradient information is not available. The key insights behind BO is the use of acquisition function, which balances exploration and exploitation, to guide the search for an optimum. In our AISTATS 2024 work "Collaborative and Explainable Bayesian Optimisation" , we propose a new BO framework that learn expert's domain knowledge in the form of a preference function. This preference function is itself another Gaussian process, which by combining with the data-driven acquisition function, allows us to combine the expert's uncertainty into the search procedure, facilitating faster convergence to an optimum. We demonstrate our method's efficacy through human-AI teaming experiments in lithium-ion battery design, highlighting substantial improvements over conventional methods.

Uncertainty in evaluations

After model training, it is crucial to evaluate their performance to ensure they meet our expectations. In addition to classical approaches like analysing confusion matrices or accuracies, significant efforts have been made to unravel the blackbox model using explanation methods. One such method is Shapley values, which attribute the model's predictions to individual features based on certain desirable fairness axioms. However, these methods are often designed for non-probabilistic functions. In our NeurIPS 2023 spotlight work "Stochastic Shapley values for Gaussian processes", we argue that the uncertainty in model predictions also plays a critical role in evaluating a model's reliability. Therefore, it is essential to develop a principled approach for uncertainty-aware feature attribution. We extend the Shapley value framework to stochastic cooperative games and propose a novel method called **Stochastic SHAP**. Additionally, we introduce a Gaussian process-specific method, GP-SHAP, which utilises the full analytical covariance structure presents in GPs to arrive at explanations that are also Gaussian random variables with fully tractable covariance structure. This covariance structure allows us to study statistical properties of the stochastic explanations, offering further insights into the blackbox function. GP-SHAP has also been applied to explain Bayesian optimisation in our AISTATS 2024 work "Collaborative and Explainable Bayesian Optimisation" , where we show that the uncertainty-aware explanations can provide more reliable insights into the model's decision-making process.

Uncertainty in deployment

After model training and evaluation, the final step is to deploy the model in real-world applications. However, most off-the-shelf learning algorithms assume that training and deployment data are independently and identically distributed (IID). In practice, this assumption is often violated due to potential distribution shifts, adversarial attacks, and strategic manipulations. In our ICML 2024 spotlight work, "Domain Generalisation via Imprecise Learning", we argue that without the IID assumption, it is no longer feasible to justify any “best” generalisation strategy. Instead, this becomes a context-specific choice problem. In other words, resolving this generalisation uncertainty requires subjective choices that should be made by the model consumer at deployment time, not by the model developer during training. To operationalise this new perspective, we propose an Imprecise Learning framework. This framework includes an optimisation process that allows learners to remain imprecise during training, avoiding commitment to a specific generalisation notion. Additionally, it provides a model framework that enables operators to define their preferred generalisation strategy at deployment without the need for model retraining.

Explanation in ML

Machine learning has achieved remarkable success in delivering accurate predictions and impressive generation capabilities. However, the opacity of these black-box models has limited their adoption in high-stakes applications. Explanation algorithms like LIME and SHAP have been developed to elucidate these black-box models by creating more understandable surrogate models that clarify the decision-making processes of complex systems. In addition to model explanations, human explanations can act as a form of prior knowledge, enhancing model performance by aligning it with expert insights. My interest in this area focuses on developing novel algorithms that utilise explanations to enable more efficient and reliable human-AI collaboration.

Constructing Explanations

Location explanation methods such as LIME and SHAP are popular because of their model-agnostic and axiomatic properties. As long as one can evaluate the function, one can deploy these methods to explain the model's predictions. However, we can often get further performance improvement (statistical or/and computational) by exploiting the model's structure. For example, when using tree-based models, due to the decompositional properties of random forest, we can significantly reduce the computational cost of SHAP computation. In our NeurIPS 2022 work "RKHS-SHAP: Shapley Values for Kernel Methods", we propose the first SHAP-based algorithm for kernel methods. By utilising the reproducing kernel Hilbert space (RKHS) structure and conditional mean embedding, we can estimate the conditional value functions without relying on sampling or density estimation methods. This approach offers a significant advantage over other methods aim to estimate observational Shapley values. Another popular function class that utilises kernels is Gaussian process models. In our NeurIPS 2023 spotlight work "Stochastic Shapley values for Gaussian processes", we extend RKHS-SHAP framework to Gaussian processes, allowing to estimate analytically both the mean and covariance of the conditional payoff functions, and subsequently leading to explanations that are now Gaussian random variables with analytical covariances. This leads to the first SHAP-based GP explanation algorithm, which we termed GP-SHAP. Besides specific model class, we also studied the problem of explaining general preference models, and proposed Pref-SHAP in our NeurIPS 2022 work "Explaining Preference with Shapley Values" to address this problem.

Utilising Explanations

Explanations can also be treated as side information to facilitate learning. In our NeurIPS 2022 work "RKHS-SHAP: Shapley Values for Kernel Methods", we proposed an attribution prior based on Shapley functionals in the RKHS, which allows us to regulate the amount of contribution certain features have to the learned function. This method are then used in fair learning to minimise the amount of contribution a sensitive feature has to the model. Explanations can also be used to improve human-AI collaboration. In our AISTATS 2024 work "Collaborative and Explainable Bayesian Optimisation" , we allow the Bayesian optimisation algorithm to suggest a pair of candidates each time, and let the human expert to choose the one for evaluation based on their domain knowledge. We showed that by explaining the acquisition function, thus letting the expert understands better why such candidate is selected, the human-AI teaming can achieve faster convergence than without explanations. Explanations can also act as an incentivising tool for changing user's behaviour. In our AAAI 2024 oral work "Causal Strategic Learning with Competitive Selection", we studied by revealing the predictive model used for selection (e.g. hiring) can incentivise strategic agents to game the system. To combat that, we proposed solutions for decision makers to be robust against gaming agents even when the model is revealed.

Preference Modelling

Recent advancements in generative AI can be largely attributed to the successful integration of human preferences into correcting model behaviours that are otherwise difficult to quantify using purely observational data. While preference learning is a classical topic in statistics (e.g. Bradley-Terry models, Thurstone model), recent breakthroughs in machine learning have opened up various new research avenues. My research in this direction aims to develop novel preference representation learning algorithms that can effectively model human preferences using various forms of feedback data, including rankings, ratings, pairwise comparisons, reviews, and choice data. Another interest of mine lies in integrating such representations into the decision-making processes of AI models, enabling them to make decisions that are more aligned with human preferences.

Preference learning

Preference learning is a classical topic in statistics, in which the goal is to learn a model that can predict the preferences of a user over a set of items. Rank prediction is a common subtask in preference learning which can be briefly categorised into two types: probabilistic approaches based on likelihood such as Bradley Terry and Plucket-Luce models, or spectral approaches that apply spectral decomposition methods to the comparison matrices directly, such as the PageRank, SerialRank, SVDRank etc. In our ECML-PKDD 2022 work "Spectral Ranking with Covariates", we extended several classical spectral ranking algorithms to incorporate covaraites, which allows us now to predict rankings of new items albeit no prior comparisons have been made. If preference data is coming in the form of ratings, in our IEEE 2021 work "Kernel-based Graph Learning from smooth signals: A Functional Viewpoint", we modelled the rating feedbacks as smooth graph signals generated from a latent comparison graph, and used kernel-based graph learning techniques to recover this latent graph structure. We also challenged the classical assumption of preferences being transitive, and proposed a more realistic non-parametric model to capture any form of preferences using our Generalised Preferential Gaussian Processes in our AISTATS 2022 work Learning Inconsistent Preferences with Gaussian processes.

Preference utilisation

Once preference are learnt, we can utilise them in various applications to improve downstream performances by injecting hard to formulate human knowledge into the model. In our AISTATS 2024 work "Collaborative and Explainable Bayesian Optimisation" , we demonstrated how preference model can be used as a prior to modify the acquisition procedure in Bayesian Optimisation to align more with expert's domain knowledge. This avoids the need for experts to specify precise priors over high dimensional input space that is known to be difficult for human to specify.