AISTATS 2025 Papers

Random feature maps are used to decrease the computational cost of kernel machines in large-scale problems. The Mondrian kernel is one such example of a fast random feature approximation of the Laplace kernel, generated by a computationally efficient hierarchical random partition of the input space known as the Mondrian process. In this work, we study a variation of this random feature map by applying a uniform random rotation to the input space before running the Mondrian process to approximate a kernel that is invariant under rotations. We obtain a closed-form expression for the isotropic kernel that is approximated, as well as a uniform convergence rate of the uniformly rotated Mondrian kernel to this limit. To this end, we utilize techniques from the theory of stationary random tessellations in stochastic geometry and prove a new result on the geometry of the typical cell of the superposition of uniformly rotated Mondrian tessellations. Finally, we test the empirical performance of this random feature map on both synthetic and real-world datasets, demonstrating its improved performance over the Mondrian kernel on a dataset that is debiased from the standard coordinate axes.

Recent studies have highlighted the benefits of generating multiple synthetic datasets for supervised learning, from increased accuracy to more effective model selection and uncertainty estimation. These benefits have clear empirical support, but the theoretical understanding of them is currently very light. We seek to increase the theoretical understanding by deriving bias-variance decompositions for several settings of using multiple synthetic datasets, including differentially private synthetic data. Our theory yields a simple rule of thumb to select the appropriate number of synthetic datasets in the case of mean-squared error and Brier score. We investigate how our theory works in practice with several real datasets, downstream predictors and error metrics. As our theory predicts, multiple synthetic datasets often improve accuracy, while a single large synthetic dataset gives at best minimal improvement, showing that our insights are practically relevant.

We introduce MODEL SELECTOR, a framework for label-efficient selection of pretrained classifiers. Given a pool of unlabeled target data, MODEL SELECTOR samples a small subset of highly informative examples for labeling, in order to efficiently identify the best pretrained model for deployment on this target dataset. Through extensive experiments, we demonstrate that MODEL SELECTOR drastically reduces the need for labeled data while consistently picking the best or near-best performing model. Across 18 model collections on 16 different datasets, comprising over 1,500 pretrained models, MODEL SELECTOR reduces the labeling cost by up to 94.15% to identify the best model compared to the cost of the strongest baseline. Our results further highlight the robustness of MODEL SELECTOR in model selection, as it reduces the labeling cost by up to 72.41% when selecting a near-best model, whose accuracy is only within 1% of the best model.

The special role of epistemic uncertainty in Machine Learning is now well recognised, and an increasing amount of research is focused on methods for dealing specifically with such a lack of knowledge. Yet, most often, a probabilistic representation is considered for both aleatoric and epistemic uncertainties, hence creating challenges in applications where decoupling these two types of uncertainty is necessary. In this work, we show that an alternative representation of epistemic uncertainty, based on possibility theory, maintains many of the convenient features of standard Bayesian inference while displaying specific behaviours and properties that closely match the ones of an intuitive notion of information. Our main contributions are: i) a general framework for jointly representing epistemic and aleatoric uncertainties, ii) a Bernstein-von Mises theorem for the analogue of Bayes' rule in possibility theory, iii) a version of the law of large numbers and of the central limit theorem for the associated variables, and iv) an analysis of the properties of the possibilistic maximum a posteriori. These results highlight that a dedicated and principled representation of epistemic uncertainty, that is compatible with standard Bayesian inference and preserves many of its strengths, is attainable.

Differential privacy (DP) provides robust privacy guarantees for statistical inference, but this can lead to unreliable results and biases in downstream applications. While several noise-aware approaches have been proposed which integrate DP perturbation into the inference, they are limited to specific types of simple probabilistic models. In this work, we propose a novel method for noise-aware approximate Bayesian inference based on stochastic gradient variational inference which can also be applied to high-dimensional and non-conjugate models. We also propose a more accurate evaluation method for noise-aware posteriors. Empirically, our inference method has similar performance to existing methods in the domain where they are applicable. Outside this domain, we obtain accurate coverages on high-dimensional Bayesian linear regression and well-calibrated predictive probabilities on Bayesian logistic regression with the UCI Adult dataset.

Deferring systems extend supervised Machine Learning (ML) models with the possibility to defer predictions to human experts. However, evaluating the impact of a deferring strategy on system accuracy is still an overlooked area. This paper fills this gap by evaluating deferring systems through a causal lens. We link the potential outcomes framework for causal inference with deferring systems, which allows to identify the causal impact of the deferring strategy on predictive accuracy. We distinguish two scenarios. In the first one, we have access to both the human and ML model predictions for the deferred instances. Here, we can identify the individual causal effects for deferred instances and the aggregates of them. In the second one, only human predictions are available for the deferred instances. Here, we can resort to regression discontinuity design to estimate a local causal effect. We evaluate our approach on synthetic and real datasets for seven deferring systems from the literature.

The Momentum Schrödinger Bridge (mSB) (Chen et al., 2023c) has emerged as a leading method for accelerating generative diffusion processes and reducing transport costs. However, the lack of simulation-free properties inevitably results in high training costs and affects scalability. To obtain a trade-off between transport properties and scalability, we introduce variational Schr\"odinger momentum diffusion (VSMD), which employs linearized forward score functions (variational scores) to eliminate the dependence on simulated forward trajectories. Our approach leverages a multivariate diffusion process with adaptively transport-optimized variational scores. Additionally, we apply a critical-damping transform to stabilize training by removing the need for score estimations for both velocity and samples. Theoretically, we prove the convergence of samples generated with optimal variational scores and momentum diffusion. Empirical results demonstrate that VSMD efficiently generates anisotropic shapes while maintaining transport efficacy, outperforming overdamped alternatives, and avoiding complex denoising processes. Our approach also scales effectively to real-world data, achieving competitive results in time series and image generation, both in unconditional and conditional settings.

Optimization problems with norm-bounding constraints appear in various applications, from portfolio optimization to machine learning, feature selection, and beyond. A widely used variant of these problems relaxes the norm-bounding constraint through Lagrangian relaxation and moves it to the objective function as a form of penalty or regularization term. A challenging class of these models uses the zero-norm function to induce sparsity in statistical parameter estimation models. Most existing exact solution methods for these problems use additional binary variables together with artificial bounds on variables to formulate them as a mixed-integer program in a higher dimension, which is then solved by off-the-shelf solvers. Other exact methods utilize specific structural properties of the objective function to solve certain variants of these problems, making them non-generalizable to other problems with different structures. An alternative approach employs nonconvex penalties with desirable statistical properties, which are solved using heuristic or local methods due to the structural complexity of those terms. In this paper, we develop a novel graph-based method to globally solve optimization problems that contain a generalization of norm-bounding constraints. This includes standard $\ell_p$-norms for $p \in [0, \infty)$ as well as nonconvex penalty terms, such as SCAD and MCP, as special cases. Our method uses decision diagrams to build strong convex relaxations for these constraints in the original space of variables without the need to introduce additional auxiliary variables or impose artificial variable bounds. We show that the resulting convexification method, when incorporated into a spatial branch-and-cut framework, converges to the global optimal value of the problem. To demonstrate the capabilities of the proposed framework, we conduct preliminary computational experiments on benchmark sparse linear regression problems with challenging nonconvex penalty terms that cannot be modeled or solved by existing global solvers.

Latent variable collaborative filtering methods have been a standard approach to modelling user-click interactions due to their simplicity and effectiveness. However, there is limited work on analyzing the mathematical properties of these methods in particular on preventing the overfitting towards the identity, and such methods typically utilize loss functions that overlook the geometry between items. In this work, we introduce a notion of generalization gap in collaborative filtering and analyze this with respect to latent collaborative filtering models. We present a geometric upper bound that gives rise to loss functions, and a way to meaningfully utilize the geometry of the items to improve recommendations. We show how these losses can be minimized and gives the recipe to a new latent collaborative filtering algorithm, which we refer to as GeoCF, due to the geometric nature of our results. We then show experimentally that our proposed GeoCF algorithm can outperform other all existing methods on the Movielens20M and Netflix datasets, as well as two large-scale internal datasets. In summary, our work proposes a theoretically sound method which paves a way to better understand generalization of collaborative filtering at large.

In this article, bipartite ranking, a statistical learning problem involved in many applications and widely studied in the passive context, is approached in a much more general active setting than the discrete one previously considered in the literature. While the latter assumes that the conditional distribution is piece wise constant, the framework we develop permits in contrast to deal with continuous conditional distributions, provided that they fulfill a Hölder smoothness constraint. We first show that a naive approach based on discretisation at a uniform level, fixed a priori and consisting in applying next the active strategy designed for the discrete setting generally fails. Instead, we propose a novel algorithm, referred to as smooth-rank and designed for the continuous setting, which aims to minimise the distance between the ROC curve of the estimated ranking rule and the optimal one w.r.t. the $\sup$ norm. We show that, for a fixed confidence level $\epsilon>0$ and probability $\delta\in (0,1)$, smooth-rank is PAC$(\epsilon,\delta)$. In addition, we provide a problem dependent upper bound on the expected sampling time of smooth-rank and establish a problem dependent lower bound on the expected sampling time of any PAC$(\epsilon,\delta)$ algorithm. Beyond the theoretical analysis carried out, numerical results are presented, providing solid empirical evidence of the performance of the algorithm proposed, which compares favorably with alternative approaches.

Conditional simulation is a fundamental task in statistical modeling: Generate samples from the conditionals given finitely many data points from a joint distribution. One promising approach is to construct conditional Brenier maps, where the components of the map pushforward a reference distribution to conditionals of the target. While many estimators exist, few, if any, come with statistical or algorithmic guarantees. To this end, we propose a non-parametric estimator for conditional Brenier maps based on the computational scalability of \emph{entropic} optimal transport. Our estimator leverages a result of Carlier et al., (2010), which shows that optimal transport maps under a rescaled quadratic cost asymptotically converge to conditional Brenier maps; our estimator is precisely the entropic analogues of these converging maps. We provide heuristic justifications for how to choose the scaling parameter in the cost as a function of the number of samples by fully characterizing the Gaussian setting. We conclude by comparing the performance of the estimator to other machine learning and non-parametric approaches on benchmark datasets and Bayesian inference problems.

We study model-based reinforcement learning with non-linear function approximation where the transition function of the underlying Markov decision process (MDP) is given by a multinomial logit (MNL) model. We develop a provably efficient discounted value iteration-based algorithm that works for both infinite-horizon average-reward and discounted-reward settings. For average-reward communicating MDPs, the algorithm guarantees a regret upper bound of $\tilde{\mathcal{O}}(dD\sqrt{T})$ where $d$ is the dimension of feature mapping, $D$ is the diameter of the underlying MDP, and $T$ is the horizon. For discounted-reward MDPs, our algorithm achieves $\tilde{\mathcal{O}}(d(1-\gamma)^{-2}\sqrt{T})$ regret where $\gamma$ is the discount factor. Then we complement these upper bounds by providing several regret lower bounds. We prove a lower bound of $\Omega(d\sqrt{DT})$ for learning communicating MDPs of diameter $D$ and a lower bound of $\Omega(d(1-\gamma)^{-3/2}\sqrt{T})$ for learning discounted-reward MDPs with discount factor $\gamma$. Lastly, we show a regret lower bound of $\Omega(dH^{3/2}\sqrt{K})$ for learning $H$-horizon episodic MDPs with MNL function approximation where $K$ is the number of episodes, which improves upon the best-known lower bound for the finite-horizon setting.

Much of science involves discovering and modeling causal relationships in nature. Significant progress has been made in developing statistical methods for representing and identifying causal knowledge from data using Linear Non-Gaussian Acyclic Models (LiNGAMs). Despite successes in learning LiNGAMs across various sample settings, the optimal sample complexity for high-dimensional LiNGAMs remains unexplored. This study establishes the optimal sample complexity for learning the structure of LiNGAMs under a sub-Gaussianity assumption. Specifically, it introduces a structure recovery algorithm using distance covariance that achieves the optimal sample complexity, $n = \Theta(d_{in} \log \frac{p}{d_{in}})$, without assuming faithfulness or a known indegree. The theoretical findings and superiority of the proposed algorithm compared to existing algorithms are validated through numerical experiments and real data analysis.

The field of unpaired image-to-image translation has undergone a significant transformation with the introduction of Generative Adversarial Networks (GANs), with CycleGAN and DiscoGAN as prominent variants. While these models show impressive empirical performance, their statistical properties are under-studied. In this paper, we propose a framework for analyzing the generalization error in cross-domain deep generative models. Our findings reveal that when provided with independent and identically distributed (i.i.d.) samples from two domains, the translation error, measured under the Wasserstein-1 loss, scales as $\tilde{\mathcal{O}} \left(\min(n, m)^{-1/\max(d,\tilde{d})}\right)$, provided that the true model possesses sufficient smoothness and the network sizes are chosen appropriately. Here, $n$ and $m$ represent the sizes of the sample sets, while $d$ and $\tilde{d}$ denote the dimensions of the respective data domains. Furthermore, we highlight the importance of a cycle loss term for ensuring distributional cycle consistency. Additionally, we provide insights into the relationship between the network size and the number of data points. Notably, as the true model exhibits greater smoothness, it suffices to work with smaller networks.

In computational physics, machine learning has now emerged as a powerful complementary tool to explore efficiently candidate designs in engineering studies. Outputs in such supervised problems are signals defined on meshes, and a natural question is the extension of general scalar output regression models to such complex outputs. Changes between input geometries in terms of both size and adjacency structure in particular make this transition non-trivial. In this work, we propose an innovative strategy for Gaussian process regression where inputs are large and sparse graphs with continuous node attributes and outputs are signals defined on the nodes of the associated inputs. The methodology relies on the combination of regularized optimal transport, dimension reduction techniques, and the use of Gaussian processes indexed by graphs. In addition to enabling signal prediction, the main point of our proposal is to come with confidence intervals on node values, which is crucial for uncertainty quantification and active learning. Numerical experiments highlight the efficiency of the method to solve real problems in fluid dynamics and solid mechanics.

The widely used Plackett-Luce ranking model assumes that individuals rank items by making repeated choices from a universe of items. But in many cases the universe is too big for people to plausibly consider all options. In the choice literature, this issue has been addressed by supposing that individuals first sample a small consideration set and then choose among the considered items. However, inferring unobserved consideration sets (or item consideration probabilities) in this ``consider then choose'' setting poses significant challenges, because even simple models of consideration with strong independence assumptions are not identifiable, even if item utilities are known. We apply the consider-then-choose framework to top-$k$ rankings, where we assume rankings are constructed according to a Plackett-Luce model after sampling a consideration set. While item consideration probabilities remain non-identified in this setting, we prove that we can infer bounds on the relative values of consideration probabilities. Additionally, given a condition on the expected consideration set size and known item utilities, we derive absolute upper and lower bounds on item consideration probabilities. We also provide algorithms to tighten those bounds on consideration probabilities by propagating inferred constraints. Thus, we show that we can learn useful information about consideration probabilities despite not being able to identify them precisely. We demonstrate our methods on a ranking dataset from a psychology experiment with two different ranking tasks (one with fixed consideration sets and one with unknown consideration sets). This combination of data allows us to estimate utilities and then learn about unknown consideration probabilities using our bounds.

Time series data are crucial across diverse domains such as finance and healthcare, where accurate forecasting and decision-making rely on advanced modeling techniques. While generative models have shown great promise in capturing the intricate dynamics inherent in time series, evaluating their performance remains a major challenge. Traditional evaluation metrics fall short due to the temporal dependencies and potential high dimensionality of the features. In this paper, we propose the REcurrent NeurAL (RENAL) Goodness-of-Fit test, a novel and statistically rigorous framework for evaluating generative time series models. By leveraging recurrent neural networks, we transform the time series into conditionally independent data pairs, enabling the application of a chi-square-based goodness-of-fit test to the temporal dependencies within the data. This approach offers a robust, theoretically grounded solution for assessing the quality of generative models, particularly in settings with limited time sequences. We demonstrate the efficacy of our method across both synthetic and real-world datasets, outperforming existing methods in terms of reliability and accuracy. Our method fills a critical gap in the evaluation of time series generative models, offering a tool that is both practical and adaptable to high-stakes applications.

We propose a family of Gaussian processes (GP) for dynamical systems with linear time-invariant responses, which are nonlinear only in initial conditions. This linearity allows us to tractably quantify forecasting and representational uncertainty, simultaneously alleviating the challenge of computing the distribution of trajectories from a GP-based dynamical system and enabling a new probabilistic treatment of learning Koopman operator representations. Using a trajectory-based equivariance – which we refer to as Koopman equivariance – we obtain a GP model with enhanced generalization capabilities. To allow for large-scale regression, we equip our framework with variational inference based on suitable inducing points. Experiments demonstrate on-par and often better forecasting performance compared to kernel-based methods for learning dynamical systems.

We propose a novel approach for optimizing the graph ratio-cut by modeling the binary assignments as random variables. We provide an upper bound on the expected ratio-cut, as well as an unbiased estimate of its gradient, to learn the parameters of the assignment variables in an online setting. The clustering resulting from our probabilistic approach (PRCut) outperforms the Rayleigh quotient relaxation of the combinatorial problem, its online learning extensions, and several widely used methods. We demonstrate that the PRCut clustering closely aligns with the similarity measure and can perform as well as a supervised classifier when label-based similarities are provided. This novel approach can leverage out-of-the-box self-supervised representations to achieve competitive performance and serve as an evaluation method for the quality of these representations.