Distribution shifts are ubiquitous in real-world machine learning applications, posing a challenge to the generalization of models trained on one data distribution to another. We focus on scenarios where data distributions vary across multiple segments of the entire population and only make local assumptions about the differences between training and test (deployment) distributions within each segment. We propose a two-stage multiply robust estimation method to improve model performance on each individual segment for tabular data analysis. The method involves fitting a linear combination of the based models, learned using clusters of training data from multiple segments, followed by a refinement step for each segment. Our method is designed to be implemented with commonly used off-the-shelf machine learning models. We establish theoretical guarantees on the generalization bound of the method on the test risk. With extensive experiments on synthetic and real datasets, we demonstrate that the proposed method substantially improves over existing alternatives in prediction accuracy and robustness on both regression and classification tasks. We also assess its effectiveness on a user city prediction dataset from a large technology company.

We study computational-statistical gaps for improper learning in sparse linear regression. More specifically, given $n$ samples from a $k$-sparse linear model in dimension $d$, we ask what is the minimum sample complexity to efficiently (in time polynomial in $d$, $k$, and $n$) find a potentially dense estimate for the regression vector that achieves non-trivial prediction error on the $n$ samples. Information-theoretically this can be achieved using $\Theta(k \log (d/k))$ samples. Yet, despite its prominence in the literature, there is no polynomial-time algorithm known to achieve the same guarantees using less than $\Theta(d)$ samples without additional restrictions on the model. Similarly, existing hardness results are either restricted to the proper setting, in which the estimate must be sparse as well, or only apply to specific algorithms. We give evidence that efficient algorithms for this task require at least (roughly) $\Omega(k^2)$ samples. In particular, we show that an improper learning algorithm for sparse linear regression can be used to solve sparse PCA problems (with a negative spike) in their Wishart form, in regimes in which efficient algorithms are widely believed to require at least $\Omega(k^2)$ samples. We complement our reduction with low-degree and statistical query lower bounds for the sparse PCA problems from which we reduce. Our hardness results apply to the (correlated) random design setting in which the covariates are drawn i.i.d. from a mean-zero Gaussian distribution with unknown covariance.

In this study, we revisit the convergence of AdaGrad with momentum (covering AdaGrad as a special case) on non-convex smooth optimization problems. We consider a general noise model where the noise magnitude is controlled by the function value gap together with the gradient magnitude. This model encompasses a broad range of noises including bounded noise, sub-Gaussian noise, affine variance noise and the expected smoothness, and it has been shown to be more realistic in many practical applications. Our analysis yields a probabilistic convergence rate which, under the general noise, could reach at (\tilde{\mathcal{O}}(1/\sqrt{T})). This rate does not rely on prior knowledge of problem-parameters and could accelerate to (\tilde{\mathcal{O}}(1/T)) where (T) denotes the total number iterations, when the noise parameters related to the function value gap and noise level are sufficiently small. The convergence rate thus matches the lower rate for stochastic first-order methods over non-convex smooth landscape up to logarithm terms [Arjevani et al., 2023]. We further derive a convergence bound for AdaGrad with mometum, considering the generalized smoothness where the local smoothness is controlled by a first-order function of the gradient norm.

For a large class of feature maps we provide a tight asymptotic characterisation of the test error associated with learning the readout layer, in the high-dimensional limit where the input dimension, hidden layer widths, and number of training samples are proportionally large. This characterization is formulated in terms of the population covariance of the features. Our work is partially motivated by the problem of learning with Gaussian rainbow neural networks, namely deep non-linear fully-connected networks with random but structured weights, whose row-wise covariances are further allowed to depend on the weights of previous layers. For such networks we also derive a closed-form formula for the feature covariance in terms of the weight matrices. We further find that in some cases our results can capture feature maps learned by deep, finite-width neural networks trained under gradient descent.

Estimating the ratio of two probability densities from finitely many samples, is a central task in machine learning and statistics. In this work, we show that a large class of kernel methods for density ratio estimation suffers from error saturation, which prevents algorithms from achieving fast error convergence rates on highly regular learning problems. To resolve saturation, we introduce iterated regularization in density ratio estimation to achieve fast error rates. Our methods outperform its non-iteratively regularized versions on benchmarks for density ratio estimation as well as on large-scale evaluations for importance-weighted ensembling of deep unsupervised domain adaptation models.

Classification models based on deep neural networks (DNNs) must be calibrated to measure the reliability of predictions. Some recent calibration methods have employed a probabilistic model on the probability simplex. However, these calibration methods cannot preserve the accuracy of pre-trained models, even those with a high classification accuracy. We propose an accuracy-preserving calibration method using the Concrete distribution as the probabilistic model on the probability simplex. We theoretically prove that a DNN model trained on cross-entropy loss has optimality as the parameter of the Concrete distribution. We also propose an efficient method that synthetically generates samples for training probabilistic models on the probability simplex. We demonstrate that the proposed method can outperform previous methods in accuracy-preserving calibration tasks using benchmarks.

The proposed solver is based on the Ohzeki method that combines Markov-chain Monte-Carlo (MCMC) and gradient descent, and its step sizes are trained by minimizing a loss function. In the training process, we propose a sampling-based gradient estimation that substitutes auto-differentiation with a variance estimation, thereby circumventing the failure of back propagation due to the non-differentiability of MCMC. The numerical results for a few COPs demonstrated that the proposed solver significantly accelerated the convergence speed compared with the original Ohzeki method. Combinatorial optimization problems (COPs) comprising discrete variables are considered hard to solve exactly in polynomial time, which relates to the well-known P vs. NP problem. Along with deterministic approximation algorithms, samplers such as Markovchain Monte-Carlo (MCMC) have been applied to COPs. However, the convergence time for obtaining reasonable approximate solutions is long.

Clustering is one of the most fundamental tools in data science and machine learning, and k-means clustering is one of the most common such methods. There is a variety of approximate algorithms for the k-means problem, but computing the globally optimal solution is in general NP-hard. In this paper we consider the k-means problem for instances with low dimensional data and formulate it as a structured concave assignment problem. This allows us to exploit the low dimensional structure and solve the problem to global optimality within reasonable time for large data sets with several clusters. The method builds on iteratively solving a small concave problem and a large linear programming problem. This gives a sequence of feasible solutions along with bounds which we show converges to zero optimality gap. The paper combines methods from global optimization theory to accelerate the procedure, and we provide numerical results on their performance.

Diarrhetic Shellfish Poisoning (DSP) is a global health threat arising from shellfish contaminated with toxins produced by dinoflagellates. The condition, with its widespread incidence, high morbidity rate, and persistent shellfish toxicity, poses risks to public health and the shellfish industry. High biomass of toxin-producing algae such as DSP are known as Harmful Algal Blooms (HABs). Monitoring and forecasting systems are crucial for mitigating HABs impact. Predicting harmful algal blooms involves a time-series-based problem with a strong historical seasonal component, however, recent anomalies due to changes in meteorological and oceanographic events have been observed. Stream Learning stands out as one of the most promising approaches for addressing time-series-based problems with concept drifts. However, its efficacy in predicting HABs remains unproven and needs to be tested in comparison with Batch Learning. Historical data availability is a critical point in developing predictive systems. In oceanography, the available data collection can have some constrains and limitations, which has led to exploring new tools to obtain more exhaustive time series. In this study, a machine learning workflow for predicting the number of cells of a toxic dinoflagellate, Dinophysis acuminata, was developed with several key advancements. Seven machine learning algorithms were compared within two learning paradigms. Notably, the output data from CROCO, the ocean hydrodynamic model, was employed as the primary dataset, palliating the limitation of time-continuous historical data. This study highlights the value of models interpretability, fair models comparison methodology, and the incorporation of Stream Learning models. The model DoME, with an average R2 of 0.77 in the 3-day-ahead prediction, emerged as the most effective and interpretable predictor, outperforming the other algorithms.

Reconstruction attacks on machine learning (ML) models pose a strong risk of leakage of sensitive data. In specific contexts, an adversary can (almost) perfectly reconstruct training data samples from a trained model using the model's gradients. When training ML models with differential privacy (DP), formal upper bounds on the success of such reconstruction attacks can be provided. So far, these bounds have been formulated under worst-case assumptions that might not hold high realistic practicality. In this work, we provide formal upper bounds on reconstruction success under realistic adversarial settings against ML models trained with DP and support these bounds with empirical results. With this, we show that in realistic scenarios, (a) the expected reconstruction success can be bounded appropriately in different contexts and by different metrics, which (b) allows for a more educated choice of a privacy parameter.