We develop a theory of transfer learning in infinitely wide neural networks where both the pretraining (source) and downstream (target) task can operate in a feature learning regime. We analyze both the Bayesian framework, where learning is described by a posterior distribution over the weights, and gradient flow training of randomly initialized networks trained with weight decay. Both settings track how representations evolve in both source and target tasks. The summary statistics of these theories are adapted feature kernels which, after transfer learning, depend on data and labels from both source and target tasks. Reuse of features during transfer learning is controlled by an elastic weight coupling which controls the reliance of the network on features learned during training on the source task. We apply our theory to linear and polynomial regression tasks as well as real datasets. Our theory and experiments reveal interesting interplays between elastic weight coupling, feature learning strength, dataset size, and source and target task alignment on the utility of transfer learning.

We study the Pareto Set Identification (PSI) problem in a structured multi-output linear bandit model. In this setting, each arm is associated a feature vector belonging to $\mathbb{R}^h$, and its mean vector in $\mathbb{R}^d$ linearly depends on this feature vector through a common unknown matrix $Θ\in \mathbb{R}^{h \times d}$. The goal is to identify the set of non-dominated arms by adaptively collecting samples from the arms. We introduce and analyze the first optimal design-based algorithms for PSI, providing nearly optimal guarantees in both the fixed-budget and the fixed-confidence settings. Notably, we show that the difficulty of these tasks mainly depends on the sub-optimality gaps of $h$ arms only. Our theoretical results are supported by an extensive benchmark on synthetic and real-world datasets.

This paper introduces KZImputer, a novel adaptive imputation method for univariate time series designed for short to medium-sized missed points (gaps) (1-5 points and beyond) with tailored strategies for segments at the start, middle, or end of the series. KZImputer employs a hybrid strategy to handle various missing data scenarios. Its core mechanism differentiates between gaps at the beginning, middle, or end of the series, applying tailored techniques at each position to optimize imputation accuracy. The method leverages linear interpolation and localized statistical measures, adapting to the characteristics of the surrounding data and the gap size. The performance of KZImputer has been systematically evaluated against established imputation techniques, demonstrating its potential to enhance data quality for subsequent time series analysis. This paper describes the KZImputer methodology in detail and discusses its effectiveness in improving the integrity of time series data. Empirical analysis demonstrates that KZImputer achieves particularly strong performance for datasets with high missingness rates (around 50% or more), maintaining stable and competitive results across statistical and signal-reconstruction metrics. The method proves especially effective in high-sparsity regimes, where traditional approaches typically experience accuracy degradation.

Causal forest methods are powerful tools in causal inference. Similar to traditional random forest in machine learning, causal forest independently considers each causal tree. However, this independence consideration increases the likelihood that classification errors in one tree are repeated in others, potentially leading to significant bias in causal e ect estimation. In this paper, we propose a novel approach that establishes connections between causal trees through the Limit Inferior Leaf Interval (LILI) clustering algorithm. LILIs are constructed based on the leaves of all causal trees, emphasizing the similarity of dataset confounders. When two instances with di erent treatments are grouped into the same leaf across a su cient number of causal trees, they are treated as counterfactual outcomes of each other. Through this clustering mechanism, LILI clustering reduces bias present in traditional causal tree methods and enhances the prediction accuracy for the average treatment e ect (ATE). By integrating LILIs into a causal forest, we develop an e cient causal inference method. Moreover, we explore several key properties of LILI by relating it to the concepts of limit inferior and limit superior in the set theory. Theoretical analysis rigorously proves the convergence of the estimated ATE using LILI clustering. Empirically, extensive comparative experiments demonstrate the superior performance of LILI clustering.

Rare diseases affect an estimated 300-400 million people worldwide, yet individual conditions often remain poorly characterized and difficult to diagnose due to their low prevalence and limited clinician familiarity. While computational phenotyping algorithms show promise for automating rare disease detection, their development is hindered by the scarcity of labeled data and biases in existing label sources. Gold-standard labels from registries and expert chart reviews are highly accurate but constrained by selection bias and the cost of manual review. In contrast, labels derived from electronic health records (EHRs) cover a broader range of patients but can introduce substantial noise. To address these challenges, we propose a weakly supervised, transformer-based framework that combines a small set of gold-standard labels with a large volume of iteratively updated silver-standard labels derived from EHR data. This hybrid approach enables the training of a highly accurate and generalizable phenotyping model that scales rare disease detection beyond the scope of individual clinical expertise. Our method is initialized by learning embeddings of medical concepts based on their semantic meaning or co-occurrence patterns in EHRs, which are then refined and aggregated into patient-level representations via a multi-layer transformer architecture. Using two rare pulmonary diseases as a case study, we validate our model on EHR data from Boston Children's Hospital. Our framework demonstrates notable improvements in phenotype classification, identification of clinically meaningful subphenotypes through patient clustering, and prediction of disease progression compared to baseline methods. These results highlight the potential of our approach to enable scalable identification and stratification of rare disease patients for clinical care and research applications.

In decision-making under uncertainty, Contextual Robust Optimization (CRO) provides reliability by minimizing the worst-case decision loss over a prediction set, hedging against label variability. While recent advances use conformal prediction to construct prediction sets for machine learning models, the downstream decisions critically depend on model selection. This paper introduces novel model selection frameworks for CRO that unify robustness control with decision risk minimization. We first propose Conformalized Robust Optimization with Model Selection (CROMS), which automatically selects models to approximately minimize the average decision risk in CRO solutions. We develop two algorithms: E-CROMS, which is computationally efficient, and F-CROMS, which enjoys a marginal robustness guarantee in finite samples. Further, we introduce Conformalized Robust Optimization with Individualized Model Selection (CROiMS), which performs individualized model selection by minimizing the conditional decision risk given the covariate of test data. This framework advances conformal prediction methodology by enabling covariate-aware model selection. Theoretically, CROiMS achieves asymptotic conditional robustness and decision efficiency under mild assumptions. Numerical results demonstrate significant improvements in decision efficiency and robustness across diverse synthetic and real-world applications, outperforming baseline approaches.

First-order methods in convex optimization offer low per-iteration cost but often suffer from slow convergence, while second-order methods achieve fast local convergence at the expense of costly Hessian inversions. In this paper, we highlight a middle ground: minimizing a quadratic majorant with fixed curvature at each iteration. This strategy strikes a balance between per-iteration cost and convergence speed, and crucially allows the reuse of matrix decompositions, such as Cholesky or spectral decompositions, across iterations and varying regularization parameters. We introduce the Quadratic Majorization Minimization with Extrapolation (QMME) framework and establish its sequential convergence properties under standard assumptions. The new perspective of our analysis is to center the arguments around the induced norm of the curvature matrix $H$. To demonstrate practical advantages, we apply QMME to large-scale kernel regularized learning problems. In particular, we propose a novel Sylvester equation modelling technique for kernel multinomial regression. In Julia-based experiments, QMME compares favorably against various established first- and second-order methods. Furthermore, we demonstrate that our algorithms complement existing kernel approximation techniques through more efficiently handling sketching matrices with large projection dimensions. Our numerical experiments and real data analysis are available and fully reproducible at https://github.com/qhengncsu/QMME.jl.

We propose a new neural network model, 01Neuro, built on indicator activation neurons. Its boosted variant possesses two key statistical properties: (1) Sure Convergence, where model optimization can be achieved with high probability given sufficient computational resources; and (2) Constructive Universal Approximation: In the infinite sample setting, the model can approximate any finite sum of measurable functions, each depending on only k out of p input features, provided the architecture is properly tuned. Unlike most universal approximation results that are agnostic to training procedures, our guarantees are directly tied to the model's explicit construction and optimization algorithm. To improve prediction stability, we integrate stochastic training and bagging into the boosted 01Neuro framework. Empirical evaluations on simulated and real-world tabular datasets with small to medium sample sizes highlight its strengths: effective approximation of interaction components (multiplicative terms), stable prediction performance (comparable to Random Forests), robustness to many noisy features, and insensitivity to feature scaling. A major limitation of the current implementation of boosted 01Neuro is its higher computational cost, which is approximately 5 to 30 times that of Random Forests and XGBoost.

Theoretical works on supervised transfer learning (STL) -- where the learner has access to labeled samples from both source and target distributions -- have for the most part focused on statistical aspects of the problem, while efficient optimization has received less attention. We consider the problem of designing an SGD procedure for STL that alternates sampling between source and target data, while maintaining statistical transfer guarantees without prior knowledge of the quality of the source data. A main algorithmic difficulty is in understanding how to design such an adaptive sub-sampling mechanism at each SGD step, to automatically gain from the source when it is informative, or bias towards the target and avoid negative transfer when the source is less informative. We show that, such a mixed-sample SGD procedure is feasible for general prediction tasks with convex losses, rooted in tracking an abstract sequence of constrained convex programs that serve to maintain the desired transfer guarantees. We instantiate these results in the concrete setting of linear regression with square loss, and show that the procedure converges, with $1/\sqrt{T}$ rate, to a solution whose statistical performance on the target is adaptive to the a priori unknown quality of the source. Experiments with synthetic and real datasets support the theory.

Modern regression problems often involve high-dimensional data and a careful tuning of the regularization hyperparameters is crucial to avoid overly complex models that may overfit the training data while guaranteeing desirable properties like effective variable selection. We study the recently introduced direction of tuning regularization hyperparameters in linear regression across multiple related tasks. We obtain distribution-dependent bounds on the generalization error for the validation loss when tuning the L1 and L2 coefficients, including ridge, lasso and the elastic net. In contrast, prior work develops bounds that apply uniformly to all distributions, but such bounds necessarily degrade with feature dimension, d. While these bounds are shown to be tight for worst-case distributions, our bounds improve with the "niceness" of the data distribution. Concretely, we show that under additional assumptions that instances within each task are i.i.d. draws from broad well-studied classes of distributions including sub-Gaussians, our generalization bounds do not get worse with increasing d, and are much sharper than prior work for very large d. We also extend our results to a generalization of ridge regression, where we achieve tighter bounds that take into account an estimate of the mean of the ground truth distribution.