Feature importance measures are widely studied and are essential for understanding model behavior, guiding feature selection, and enhancing interpretability. However, many machine learning fitted models involve complex, higher-order interactions between features. Existing feature importance metrics fail to capture these higher-order effects while existing interaction metrics often suffer from limited applicability or excessive computation; no methods exist to conduct statistical inference for feature interactions. To bridge this gap, we first propose a new model-agnostic metric, interaction Leave-One-Covariate-Out iLOCO, for measuring the importance of higher-order feature interactions. Next, we leverage recent advances in LOCO inference to develop distribution-free and assumption-light confidence intervals for our iLOCO metric. To address computational challenges, we also introduce an ensemble learning method for calculating the iLOCO metric and confidence intervals that we show is both computationally and statistically efficient. We validate our iLOCO metric and our confidence intervals on both synthetic and real data sets, showing that our approach outperforms existing methods and provides the first inferential approach to detecting feature interactions.

Using observational studies to estimate treatment effects is a ubiquitous yet challenging task in many disciplines, such as medicine [Hernán and Robins, 2006] or social sciences [Athey and Imbens, 2017]. Whereas there exists a rich literature of methods for treatment effect estimation in the observational setting [Bang and Robins, 2005, Wager and Athey, 2018, Chernozhukov et al., 2018], all methods have in common that before a causal effect can be estimated, often untestable conditions need to hold. One such condition is that we assume there is no unmeasured confounding, meaning that there are no unobserved factors that have both an influence on the treatment and on the outcome of interest that are not accounted for by the method. If unmeasured confounders are present, our causal effect estimates are likely to be biased and inconsistent [Greenland et al., 1999]. This can have serious downstream consequences such as unknowingly recommending a non-effective or, even worse, potentially harmful treatment policy. Unfortunately, without making further assumptions, it is in general impossible to verify all assumptions needed to identify treatment effects from observational data. In this work, we investigate a novel strategy for falsifying unconfoundedness. Specifically, we focus on the common scenario where observational datasets are collected from different heterogeneous sources, which we refer to as environments.

In Earth sciences, unobserved factors exhibit non-stationary spatial distributions, causing the relationships between features and targets to display spatial heterogeneity. In geographic machine learning tasks, conventional statistical learning methods often struggle to capture spatial heterogeneity, leading to unsatisfactory prediction accuracy and unreliable interpretability. While approaches like Geographically Weighted Regression (GWR) capture local variations, they fall short of uncovering global patterns and tracking the continuous evolution of spatial heterogeneity. Motivated by this limitation, we propose a novel perspective - that is, simultaneously modeling common features across different locations alongside spatial differences using deep neural networks. The proposed method is a dual-branch neural network with an encoder-decoder structure. In the encoding stage, the method aggregates node information in a spatiotemporal conditional graph using GCN and LSTM, encoding location-specific spatiotemporal heterogeneity as an implicit conditional vector. Additionally, a self-attention-based encoder is used to extract location-invariant common features from the data. In the decoding stage, the approach employs a conditional generation strategy that predicts response variables and interpretative weights based on data features under spatiotemporal conditions. The approach is validated by predicting vegetation gross primary productivity (GPP) using global climate and land cover data from 2001 to 2020. Trained on 50 million samples and tested on 2.8 million, the proposed model achieves an RMSE of 0.836, outperforming LightGBM (1.063) and TabNet (0.944). Visualization analyses indicate that our method can reveal the distribution differences of the dominant factors of GPP across various times and locations.

A plethora of real-world scientific investigations is waiting to scale with the support of trustworthy predictive models that can reduce the need for costly data annotations. We focus on causal inferences on a target experiment with unlabeled factual outcomes, retrieved by a predictive model fine-tuned on a labeled similar experiment. First, we show that factual outcome estimation via Empirical Risk Minimization (ERM) may fail to yield valid causal inferences on the target population, even in a randomized controlled experiment and infinite training samples. Then, we propose to leverage the observed experimental settings during training to empower generalization to downstream interventional investigations, ``Causal Lifting'' the predictive model. We propose Deconfounded Empirical Risk Minimization (DERM), a new simple learning procedure minimizing the risk over a fictitious target population, preventing potential confounding effects. We validate our method on both synthetic and real-world scientific data. Notably, for the first time, we zero-shot generalize causal inferences on ISTAnt dataset (without annotation) by causal lifting a predictive model on our experiment variant.

In this paper, we establish non-asymptotic convergence rates in the central limit theorem for Polyak-Ruppert-averaged iterates of stochastic gradient descent (SGD). Our analysis builds on the result of the Gaussian approximation for nonlinear statistics of independent random variables of Shao and Zhang (2022). Using this result, we prove the non-asymptotic validity of the multiplier bootstrap for constructing the confidence sets for the optimal solution of an optimization problem. In particular, our approach avoids the need to approximate the limiting covariance of Polyak-Ruppert SGD iterates, which allows us to derive approximation rates in convex distance of order up to $1/\sqrt{n}$.

The problem of learning single index and multi index models has gained significant interest as a fundamental task in high-dimensional statistics. Many recent works have analysed gradient-based methods, particularly in the setting of isotropic data distributions, often in the context of neural network training. Such studies have uncovered precise characterisations of algorithmic sample complexity in terms of certain analytic properties of the target function, such as the leap, information, and generative exponents. These properties establish a quantitative separation between low and high complexity learning tasks. In this work, we show that high complexity cases are rare. Specifically, we prove that introducing a small random perturbation to the data distribution--via a random shift in the first moment--renders any Gaussian single index model as easy to learn as a linear function. We further extend this result to a class of multi index models, namely sparse Boolean functions, also known as Juntas.

Recent advances in machine learning have transformed our ability to develop high quality predictive and generative models for complex data. This work introduces a framework for developing decisionmaking algorithms, specifically for contextual bandit problems, that can take advantage of these machine learning advances. By design, we assume the algorithm developer is able to effectively apply these machine learning techniques (e.g., minimize a loss via gradient descent) and employ these methods as subroutines in our decision-making algorithm. Moreover, our theory formally connects the quality of effective (self-)supervised learning via loss minimization to the quality of decision-making. Classically, contextual Thompson sampling algorithms form a parametric model of the environment and consider the decision-maker's uncertainty as arising from unknown latent parameters of that model [51]. In this classical perspective, the primitive operations that are assumed to be feasible (at least approximately) include i) the ability to specify an informative prior for the latent parameter using domain knowledge, ii) the ability to sample from the posterior distribution of the latent parameter, and iii) the ability to update the posterior distribution as more data is collected. Unfortunately, it is well known that all three of these primitive operations are non-trivial to perform with neural networks [20, 61]. Building on our previous work [8] which focuses on multi-armed bandits without contexts, we view missing, but potentially observable, future outcomes as the source of the decision-maker's uncertainty. This perspective allows us to replace the primitive operations required in the classical view with new primitives that are more compatible with neural networks: i) the ability to effectively minimize an offline sequence prediction loss, ii) the ability to autoregressively generate from the optimized sequence model, and iii) the ability to fit a desired policy given access to a complete dataset (outcomes from all actions and decision-times).

In statistics and machine learning, when we train a fitted model on available data, we typically want to ensure that we are searching within a model class that contains at least one accurate model -- that is, we would like to ensure an upper bound on the model class risk (the lowest possible risk that can be attained by any model in the class). However, it is also of interest to establish lower bounds on the model class risk, for instance so that we can determine whether our fitted model is at least approximately optimal within the class, or, so that we can decide whether the model class is unsuitable for the particular task at hand. Particularly in the setting of interpolation learning where machine learning models are trained to reach zero error on the training data, we might ask if, at the very least, a positive lower bound on the model class risk is possible -- or are we unable to detect that "all models are wrong"? In this work, we answer these questions in a distribution-free setting by establishing a model-agnostic, fundamental hardness result for the problem of constructing a lower bound on the best test error achievable over a model class, and examine its implications on specific model classes such as tree-based methods and linear regression.

The particle filter (PF) and the ensemble Kalman filter (EnKF) are widely used for approximate inference in state-space models. From a Bayesian perspective, these algorithms represent the prior by an ensemble of particles and update it to the posterior with new observations over time. However, the PF often suffers from weight degeneracy in high-dimensional settings, whereas the EnKF relies on linear Gaussian assumptions that can introduce significant approximation errors. In this paper, we propose the Adversarial Transform Particle Filter (ATPF), a novel filtering framework that combines the strengths of the PF and the EnKF through adversarial learning. Specifically, importance sampling is used to ensure statistical consistency as in the PF, while adversarially learned transformations, such as neural networks, allow accurate posterior matching for nonlinear and non-Gaussian systems. In addition, we incorporate kernel methods to ease optimization and leverage regularization techniques based on optimal transport for better statistical properties and numerical stability. We provide theoretical guarantees, including generalization bounds for both the analysis and forecast steps of ATPF. Extensive experiments across various nonlinear and non-Gaussian scenarios demonstrate the effectiveness and practical advantages of our method.

In domain adaptation (DA), the effectiveness of deep learning-based models is often constrained by batch learning strategies that fail to fully apprehend the global statistical and geometric characteristics of data distributions. Addressing this gap, we introduce 'Global Awareness Enhanced Domain Adaptation' (GAN-DA), a novel approach that transcends traditional batch-based limitations. GAN-DA integrates a unique predefined feature representation (PFR) to facilitate the alignment of cross-domain distributions, thereby achieving a comprehensive global statistical awareness. This representation is innovatively expanded to encompass orthogonal and common feature aspects, which enhances the unification of global manifold structures and refines decision boundaries for more effective DA. Our extensive experiments, encompassing 27 diverse cross-domain image classification tasks, demonstrate GAN-DA's remarkable superiority, outperforming 24 established DA methods by a significant margin. Furthermore, our in-depth analyses shed light on the decision-making processes, revealing insights into the adaptability and efficiency of GAN-DA. This approach not only addresses the limitations of existing DA methodologies but also sets a new benchmark in the realm of domain adaptation, offering broad implications for future research and applications in this field.