AI-driven precision oncology has the transformative potential to reshape cancer treatment by leveraging the power of AI models to analyze the interaction between complex patient characteristics and their corresponding treatment outcomes. New technological platforms have facilitated the timely acquisition of multimodal data on tumor biology at an unprecedented resolution, such as single-cell multi-omics data, making this quality and quantity of data available for data-driven improved clinical decision-making. In this work, we propose a modular machine learning framework designed for personalized counterfactual cancer treatment suggestions based on an ensemble of machine learning experts trained on diverse multi-omics technologies. These specialized counterfactual experts per technology are consistently aggregated into a more powerful expert with superior performance and can provide both confidence and an explanation of its decision. The framework is tailored to address critical challenges inherent in data-driven cancer research, including the high-dimensional nature of the data, and the presence of treatment assignment bias in the retrospective observational data. The framework is showcased through comprehensive demonstrations using data from in-vitro and in-vivo treatment responses from a cohort of patients with ovarian cancer. Our method aims to empower clinicians with a reality-centric decision-support tool including probabilistic treatment suggestions with calibrated confidence and personalized explanations for tailoring treatment strategies to multi-omics characteristics of individual cancer patients.

Clustering algorithms are used extensively in data analysis for data exploration and discovery. Technological advancements lead to continually growth of data in terms of volume, dimensionality and complexity. This provides great opportunities in data analytics as the data can be interrogated for many different purposes. This however leads challenges, such as identification of relevant features for a given task. In supervised tasks, one can utilise a number of methods to optimise the input features for the task objective (e.g. classification accuracy). In unsupervised problems, such tools are not readily available, in part due to an inability to quantify feature relevance in unlabeled tasks. In this paper, we investigate the sensitivity of clustering performance noisy uncorrelated variables iteratively added to baseline datasets with well defined clusters. We show how different types of irrelevant variables can impact the outcome of a clustering result from $k$-means in different ways. We observe a resilience to very high proportions of irrelevant features for adjusted rand index (ARI) and normalised mutual information (NMI) when the irrelevant features are Gaussian distributed. For Uniformly distributed irrelevant features, we notice the resilience of ARI and NMI is dependent on the dimensionality of the data and exhibits tipping points between high scores and near zero. Our results show that the Silhouette Coefficient and the Davies-Bouldin score are the most sensitive to irrelevant added features exhibiting large changes in score for comparably low proportions of irrelevant features regardless of underlying distribution or data scaling. As such the Silhouette Coefficient and the Davies-Bouldin score are good candidates for optimising feature selection in unsupervised clustering tasks.

Distributionally robust optimization has emerged as an attractive way to train robust machine learning models, capturing data uncertainty and distribution shifts. Recent statistical analyses have proved that robust models built from Wasserstein ambiguity sets have nice generalization guarantees, breaking the curse of dimensionality. However, these results are obtained in specific cases, at the cost of approximations, or under assumptions difficult to verify in practice. In contrast, we establish, in this article, exact generalization guarantees that cover all practical cases, including any transport cost function and any loss function, potentially non-convex and nonsmooth. For instance, our result applies to deep learning, without requiring restrictive assumptions. We achieve this result through a novel proof technique that combines nonsmooth analysis rationale with classical concentration results. Our approach is general enough to extend to the recent versions of Wasserstein/Sinkhorn distributionally robust problems that involve (double) regularizations.

Interactions among large number of entities is naturally high-dimensional and incomplete (HDI) in many big data related tasks. Behavioral characteristics of users are hidden in these interactions, hence, effective representation of the HDI data is a fundamental task for understanding user behaviors. Latent factor analysis (LFA) model has proven to be effective in representing HDI data. The performance of an LFA model relies heavily on its training process, which is a non-convex optimization. It has been proven that incorporating local curvature and preprocessing gradients during its training process can lead to superior performance compared to LFA models built with first-order family methods. However, with the escalation of data volume, the feasibility of second-order algorithms encounters challenges. To address this pivotal issue, this paper proposes a mini-block diagonal hessian-free (Mini-Hes) optimization for building an LFA model. It leverages the dominant diagonal blocks in the generalized Gauss-Newton matrix based on the analysis of the Hessian matrix of LFA model and serves as an intermediary strategy bridging the gap between first-order and second-order optimization methods. Experiment results indicate that, with Mini-Hes, the LFA model outperforms several state-of-the-art models in addressing missing data estimation task on multiple real HDI datasets from recommender system. (The source code of Mini-Hes is available at https://github.com/Goallow/Mini-Hes)

This paper studies the fitting of Hessian or its inverse with stochastic Hessian-vector products. A Hessian fitting criterion, which can be used to derive most of the commonly used methods, e.g., BFGS, Gaussian-Newton, AdaGrad, etc., is used for the analysis. Our studies reveal different convergence rates for different Hessian fitting methods, e.g., sublinear rates for gradient descent in the Euclidean space and a commonly used closed-form solution, linear rates for gradient descent on the manifold of symmetric positive definite (SPL) matrices and certain Lie groups. The Hessian fitting problem is further shown to be strongly convex under mild conditions on a specific yet general enough Lie group. To confirm our analysis, these methods are tested under different settings like noisy Hessian-vector products, time varying Hessians, and low precision arithmetic. These findings are useful for stochastic second order optimizations that rely on fast, robust and accurate Hessian estimations.

Transformer-based architectures achieved breakthrough performance in natural language processing and computer vision, yet they remain inferior to simpler linear baselines in multivariate long-term forecasting. To better understand this phenomenon, we start by studying a toy linear forecasting problem for which we show that transformers are incapable of converging to their true solution despite their high expressive power. We further identify the attention of transformers as being responsible for this low generalization capacity. Building upon this insight, we propose a shallow lightweight transformer model that successfully escapes bad local minima when optimized with sharpness-aware optimization. We empirically demonstrate that this result extends to all commonly used real-world multivariate time series datasets. In particular, SAMformer surpasses the current state-of-the-art model TSMixer by 14.33% on average, while having ~4 times fewer parameters. The code is available at https://github.com/romilbert/samformer.

Online transaction fraud presents substantial challenges to businesses and consumers, risking significant financial losses. Conventional rule-based systems struggle to keep pace with evolving fraud tactics, leading to high false positive rates and missed detections. Machine learning techniques offer a promising solution by leveraging historical data to identify fraudulent patterns. This article explores using the personalised PageRank (PPR) algorithm to capture the social dynamics of fraud by analysing relationships between financial accounts. The primary objective is to compare the performance of traditional features with the addition of PPR in fraud detection models. Results indicate that integrating PPR enhances the model's predictive power, surpassing the baseline model. Additionally, the PPR feature provides unique and valuable information, evidenced by its high feature importance score. Feature stability analysis confirms consistent feature distributions across training and test datasets.

This thesis explores the generation of local explanations for already deployed machine learning models, aiming to identify optimal conditions for producing meaningful explanations considering both data and user requirements. The primary goal is to develop methods for generating explanations for any model while ensuring that these explanations remain faithful to the underlying model and comprehensible to the users. The thesis is divided into two parts. The first enhances a widely used rule-based explanation method. It then introduces a novel approach for evaluating the suitability of linear explanations to approximate a model. Additionally, it conducts a comparative experiment between two families of counterfactual explanation methods to analyze the advantages of one over the other. The second part focuses on user experiments to assess the impact of three explanation methods and two distinct representations. These experiments measure how users perceive their interaction with the model in terms of understanding and trust, depending on the explanations and representations. This research contributes to a better explanation generation, with potential implications for enhancing the transparency, trustworthiness, and usability of deployed AI systems.

Stroke is a common disabling neurological condition that affects about one-quarter of the adult population over age 25; more than half of patients still have poor outcomes, such as permanent functional dependence or even death, after the onset of acute stroke. The aim of this study is to investigate the efficacy of diffusion-weighted MRI modalities combining with structured health profile on predicting the functional outcome to facilitate early intervention. A deep fusion learning network is proposed with two-stage training: the first stage focuses on cross-modality representation learning and the second stage on classification. Supervised contrastive learning is exploited to learn discriminative features that separate the two classes of patients from embeddings of individual modalities and from the fused multimodal embedding. The network takes as the input DWI and ADC images, and structured health profile data. The outcome is the prediction of the patient needing long-term care at 3 months after the onset of stroke. Trained and evaluated with a dataset of 3297 patients, our proposed fusion model achieves 0.87, 0.80 and 80.45% for AUC, F1-score and accuracy, respectively, outperforming existing models that consolidate both imaging and structured data in the medical domain. If trained with comprehensive clinical variables, including NIHSS and comorbidities, the gain from images on making accurate prediction is not considered substantial, but significant. However, diffusion-weighted MRI can replace NIHSS to achieve comparable level of accuracy combining with other readily available clinical variables for better generalization.

While deep learning has enabled tremendous progress on text and image datasets, its superiority on tabular data is not clear. We contribute extensive benchmarks of standard and novel deep learning methods as well as tree-based models such as XGBoost and Random Forests, across a large number of datasets and hyperparameter combinations. We define a standard set of 45 datasets from varied domains with clear characteristics of tabular data and a benchmarking methodology accounting for both fitting models and finding good hyperparameters. Results show that tree-based models remain state-of-the-art on medium-sized data ( 10K samples) even without accounting for their superior speed. To understand this gap, we conduct an empirical investigation into the differing inductive biases of tree-based models and neural networks. This leads to a series of challenges which should guide researchers aiming to build tabular-specific neural network: 1. be robust to uninformative features, 2. preserve the orientation of the data, and 3. be able to easily learn irregular functions. To stimulate research on tabular architectures, we contribute a standard benchmark and raw data for baselines: every point of a 20 000 compute hours hyperparameter search for each learner.