Room 02.124, Building 5, St Priest campus Machine Learning in Montpellier, Theory & Practice Optimal Transport is a useful metric to compare probability distributions and to compute a pairing given a ground cost. Its entropic regularization variant (eOT) is crucial to have fast algorithms and reflect fuzzy/noisy matchings. This work focuses on Inverse Optimal Transport (iOT), the problem of inferring the ground cost from samples drawn from a coupling that solves an eOT problem. It is a relevant problem that can be used to infer unobserved/missing links, and to obtain meaningful information about the structure of the ground cost yielding the pairing. On one side, iOT benefits from convexity, but on the other side, being ill-posed, it requires regularization to handle the sampling noise. This work presents an in-depth theoretical study of the l1 regularization to model for instance Euclidean costs with sparse interactions between features. Specifically, we derive a sufficient condition for the robust recovery of the sparsity of the ground cost that can be seen as a far reaching generalization of the Lasso's celebrated Irrepresentability Condition. To provide additional insight into this condition, we work out in detail the Gaussian case. We show that as the entropic penalty varies, the iOT problem interpolates between a graphical Lasso and a classical Lasso, thereby establishing a connection between iOT and graph estimation, an important problem in ML. Machine Learning in Montpellier, Theory & Practice

Room 02.124, Building 5, St Priest campus Machine Learning in Montpellier, Theory & Practice Causal inference using observational data presents many statistical challenges, particularly when dealing with missing confounder data. While multiple imputation offers a potential solution, its implementation with propensity score matching requires careful consideration. In this talk, we will delve into empirical studies conducted by the LSHTM* Statistics team to explore these matters.,, Machine Learning in Montpellier, Theory & Practice

Room 02.124, Building 5, St Priest campus Machine Learning in Montpellier, Theory & Practice Overfitting on training data is typically assumed to be a bad practice. However, modern machine learning models are highly over-parameterised, to the extent that they can overfit on training data. In this talk, I will discuss new theories that debunk the myths that: (1) large models with too many parameters always overfit the training data; and (2) models that perfectly fit the training data cannot predict well on new data.I will then present our recent works on generalisation and learning dynamics of over-parameterised models, including (1) the double descent phenomenon in causal inference and (2) the neural tangent kernel approximation for semi- and self-supervised models. I will highlight how some of our results resolve conjectures in machine learning, and also provide new practical tools.,, Machine Learning in Montpellier, Theory & Practice

Room 02.124, Building 5, St Priest campus Machine Learning in Montpellier, Theory & Practice While classification datasets are composed of more and more data, the need for human expertise to label them is still present. Crowdsourcing platforms are a way to gather expert feedback at a low cost. However, the quality of these labels is not always guaranteed. In this thesis, we focus on the problem of label ambiguity in crowdsourcing. Label ambiguity has mostly two sources: the worker’s ability and the task’s difficulty. We first present a new indicator, the WAUM (Weighted Area Under the Magin), to detect ambiguous tasks given to workers. Based on the existing AUM in the classical supervised setting, this lets us explore large datasets while focusing on tasks that might require more relevant expertise or should be discarded from the actual dataset. We then present a new open-source python library, PeerAnnot, that we developed to handle crowdsourced datasets in image classification. Finally, we present a case study on the Pl@ntNet dataset, where we evaluate the current state of the platform’s label aggregation strategy and propose ways to improve it. This setting, with a large number of tasks, experts and classes, is highly challenging for current crowdsourcing aggregation strategies. We report consistently better performance against competitors and propose a new aggregation strategy that could be used in the future to improve the quality of the Pl@ntNet dataset. We also release this large dataset of expert feedback that could be used to improve the quality of the current aggregation methods and provide a new benchmark,, Machine Learning in Montpellier, Theory & Practice

Room Saltus, Maison de la Télédétection / Online Machine Learning in Montpellier, Theory & Practice Machine Learning in Montpellier, Theory & Practice

Online Machine Learning in Montpellier, Theory & Practice In many fields, including medicine, education, and public policy, researchers and practitioners are interested in determining what works for whom - in other words, identifying subgroups for whom specific interventions work particularly well. By finding these subgroups, we can more effectively focus resources and give interventions to those who might benefit the most. These individualized treatment decisions can improve outcomes, but answering these types of questions is challenging with a single dataset. Specifically, randomized controlled trials have unconfounded treatment assignment but are often too small to reliably estimate heterogeneous treatment effects, while larger observational datasets might suffer from confounding. Data integration methods can utilize the benefits of different sources of data while accounting for bias. In this talk, I first discuss non-parametric data integration approaches for combining multiple randomized controlled trials to estimate the effect of treatments conditional on observed characteristics. I explore the performance of these methods through a simulation study, and I apply the approaches to four randomized controlled trials to examine effect heterogeneity of treatments for major depressive disorder. I then discuss methods for applying these multi study treatment effect models to an external, observational target sample represented by electronic health records of a set of patients. With these methods, we can utilize individual-level data across sources to improve our ability to make intervention decisions that are tailored to individuals or communities, and we can ultimately apply our conclusions to a given target population. Machine Learning in Montpellier, Theory & Practice

Online Machine Learning in Montpellier, Theory & Practice Unmeasured confounding is a major challenge for identifying causal relationships from non-experimental data. Here, we propose a method that can address unmeasured discrete confounding. Extending recent identifiability results in deep latent variable models, we show theoretically that confounding can be detected and corrected under the assumption that the observed data is a piecewise affine transformation of a latent Gaussian mixture model and that the identity of the mixture components is confounded. We provide a flow-based algorithm to estimate this model and perform deconfounding. Experimental results on synthetic and real-world data provide support for the effectiveness of our approach. Machine Learning in Montpellier, Theory & Practice

Online Machine Learning in Montpellier, Theory & Practice Many modern causal questions ask how treatments affect complex outcomes that are measured using wearable devices and sensors. Current analysis approaches require summarizing these data into scalar statistics (e.g., the mean), but these summaries can be misleading. For example, disparate distributions can have the same means, variances, and other statistics. Researchers can overcome the loss in information by instead representing the data as distributions. We develop an interpretable method for distributional data analysis that ensures trustworthy and robust decision making: Analyzing Distributional Data via Matching After Learning to Stretch (ADD MALTS). We (i) provide analytical guarantees of the correctness of our estimation strategy,(ii) demonstrate via simulation that ADD MALTS outperforms other distributional data analysis methods at estimating treatment effects, and (iii) illustrate ADD MALTS’ ability to verify whether there is enough cohesion between treatment and control units within subpopulations to trustworthily estimate treatment effects. We demonstrate ADD MALTS’ utility by studying the effectiveness of continuous glucose monitors in mitigating diabetes risks. Machine Learning in Montpellier, Theory & Practice

Online Machine Learning in Montpellier, Theory & Practice Machine Learning in Montpellier, Theory & Practice