Bagging is a popular ensemble technique to improve the accuracy of machine learning models. It hinges on the well-established rationale that, by repeatedly retraining on resampled data, the aggregated model exhibits lower variance and hence higher stability, especially for discontinuous base learners. In this paper, we provide a new perspective on bagging: By suitably aggregating the base learners at the parametrization instead of the output level, bagging improves generalization performances exponentially, a strength that is significantly more powerful than variance reduction. More precisely, we show that for general stochastic optimization problems that suffer from slowly (i.e., polynomially) decaying generalization errors, bagging can effectively reduce these errors to an exponential decay. Moreover, this power of bagging is agnostic to the solution schemes, including common empirical risk minimization, distributionally robust optimization, and various regularizations. We demonstrate how bagging can substantially improve generalization performances in a range of examples involving heavy-tailed data that suffer from intrinsically slow rates.
Direct Preference Optimization (DPO) has recently emerged as a popular approach to improve reinforcement learning with human feedback (RLHF), leading to better techniques to fine-tune large language models (LLM). A weakness of DPO, however, lies in its lack of capability to characterize the diversity of human preferences. Inspired by Mallows' theory of preference ranking, we develop in this paper a new approach, the Mallows-DPO. A distinct feature of this approach is a dispersion index, which reflects the dispersion of human preference to prompts. We show that existing DPO models can be reduced to special cases of this dispersion index, thus unified with Mallows-DPO. More importantly, we demonstrate (empirically) how to use this dispersion index to enhance the performance of DPO in a broad array of benchmark tasks, from synthetic bandit selection to controllable generations and dialogues, while maintaining great generalization capabilities.
Offline reinforcement learning (RL) offers a promising direction for learning policies from pre-collected datasets without requiring further interactions with the environment. However, existing methods struggle to handle out-of-distribution (OOD) extrapolation errors, especially in sparse reward or scarce data settings. In this paper, we propose a novel training algorithm called Conservative Density Estimation (CDE), which addresses this challenge by explicitly imposing constraints on the state-action occupancy stationary distribution. CDE overcomes the limitations of existing approaches, such as the stationary distribution correction method, by addressing the support mismatch issue in marginal importance sampling. Our method achieves state-of-the-art performance on the D4RL benchmark. Notably, CDE consistently outperforms baselines in challenging tasks with sparse rewards or insufficient data, demonstrating the advantages of our approach in addressing the extrapolation error problem in offline RL.
Stochastic gradient descent (SGD) or stochastic approximation has been widely used in model training and stochastic optimization. While there is a huge literature on analyzing its convergence, inference on the obtained solutions from SGD has only been recently studied, yet is important due to the growing need for uncertainty quantification. We investigate two computationally cheap resampling-based methods to construct confidence intervals for SGD solutions. One uses multiple, but few, SGDs in parallel via resampling with replacement from the data, and another operates this in an online fashion. Our methods can be regarded as enhancements of established bootstrap schemes to substantially reduce the computation effort in terms of resampling requirements, while at the same time bypassing the intricate mixing conditions in existing batching methods. We achieve these via a recent so-called cheap bootstrap idea and Berry-Esseen-type bound for SGD.
Bayesian Optimization is a popular approach for optimizing expensive black-box functions. Its key idea is to use a surrogate model to approximate the objective and, importantly, quantify the associated uncertainty that allows a sequential search of query points that balance exploitation-exploration. Gaussian process (GP) has been a primary candidate for the surrogate model, thanks to its Bayesian-principled uncertainty quantification power and modeling flexibility. However, its challenges have also spurred an array of alternatives whose convergence properties could be more opaque. Motivated by these, we study in this paper an axiomatic framework that elicits the minimal requirements to guarantee black-box optimization convergence that could apply beyond GP-related methods. Moreover, we leverage the design freedom in our framework, which we call Pseudo-Bayesian Optimization, to construct empirically superior algorithms. In particular, we show how using simple local regression, and a suitable "randomized prior" construction to quantify uncertainty, not only guarantees convergence but also consistently outperforms state-of-the-art benchmarks in examples ranging from high-dimensional synthetic experiments to realistic hyperparameter tuning and robotic applications.
In data-driven optimization, sample average approximation is known to suffer from the so-called optimizer's curse that causes optimistic bias in evaluating the solution performance. This can be tackled by adding a "margin" to the estimated objective value, or via distributionally robust optimization (DRO), a fast-growing approach based on worst-case analysis, which gives a protective bound on the attained objective value. However, in all these existing approaches, a statistically guaranteed bound on the true solution performance either requires restrictive conditions and knowledge on the objective function complexity, or otherwise exhibits an over-conservative rate that depends on the distribution dimension. We argue that a special type of DRO offers strong theoretical advantages in regard to these challenges: It attains a statistical bound on the true solution performance that is the tightest possible in terms of exponential decay rate, for a wide class of objective functions that notably does not hinge on function complexity. Correspondingly, its calibration also does not require any complexity information. This DRO uses an ambiguity set based on a KL-divergence smoothed by the Wasserstein or Levy-Prokhorov distance via a suitable distance optimization. Computationally, we also show that such a DRO, and its generalized version using smoothed $f$-divergence, is not much harder than standard DRO problems using the $f$-divergence or Wasserstein distance, thus supporting the strengths of such DRO as both statistically optimal and computationally viable.
In data-driven optimization, the sample performance of the obtained decision typically incurs an optimistic bias against the true performance, a phenomenon commonly known as the Optimizer's Curse and intimately related to overfitting in machine learning. Common techniques to correct this bias, such as cross-validation, require repeatedly solving additional optimization problems and are therefore computationally expensive. We develop a general bias correction approach, building on what we call Optimizer's Information Criterion (OIC), that directly approximates the first-order bias and does not require solving any additional optimization problems. Our OIC generalizes the celebrated Akaike Information Criterion to evaluate the objective performance in data-driven optimization, which crucially involves not only model fitting but also its interplay with the downstream optimization. As such it can be used for decision selection instead of only model selection. We apply our approach to a range of data-driven optimization formulations comprising empirical and parametric models, their regularized counterparts, and furthermore contextual optimization. Finally, we provide numerical validation on the superior performance of our approach under synthetic and real-world datasets.
Uncertainty quantification (UQ) is important for reliability assessment and enhancement of machine learning models. In deep learning, uncertainties arise not only from data, but also from the training procedure that often injects substantial noises and biases. These hinder the attainment of statistical guarantees and, moreover, impose computational challenges on UQ due to the need for repeated network retraining. Building upon the recent neural tangent kernel theory, we create statistically guaranteed schemes to principally \emph{quantify}, and \emph{remove}, the procedural uncertainty of over-parameterized neural networks with very low computation effort. In particular, our approach, based on what we call a procedural-noise-correcting (PNC) predictor, removes the procedural uncertainty by using only \emph{one} auxiliary network that is trained on a suitably labeled data set, instead of many retrained networks employed in deep ensembles. Moreover, by combining our PNC predictor with suitable light-computation resampling methods, we build several approaches to construct asymptotically exact-coverage confidence intervals using as low as four trained networks without additional overheads.
Many event sequence data exhibit mutually exciting or inhibiting patterns. Reliable detection of such temporal dependency is crucial for scientific investigation. The de facto model is the Multivariate Hawkes Process (MHP), whose impact function naturally encodes a causal structure in Granger causality. However, the vast majority of existing methods use direct or nonlinear transform of standard MHP intensity with constant baseline, inconsistent with real-world data. Under irregular and unknown heterogeneous intensity, capturing temporal dependency is hard as one struggles to distinguish the effect of mutual interaction from that of intensity fluctuation. In this paper, we address the short-term temporal dependency detection issue. We show the maximum likelihood estimation (MLE) for cross-impact from MHP has an error that can not be eliminated but may be reduced by order of magnitude, using heterogeneous intensity not of the target HP but of the interacting HP. Then we proposed a robust and computationally-efficient method modified from MLE that does not rely on the prior estimation of the heterogeneous intensity and is thus applicable in a data-limited regime (e.g., few-shot, no repeated observations). Extensive experiments on various datasets show that our method outperforms existing ones by notable margins, with highlighted novel applications in neuroscience.
In data-driven stochastic optimization, model parameters of the underlying distribution need to be estimated from data in addition to the optimization task. Recent literature suggests the integration of the estimation and optimization processes, by selecting model parameters that lead to the best empirical objective performance. Such an integrated approach can be readily shown to outperform simple ``estimate then optimize" when the model is misspecified. In this paper, we argue that when the model class is rich enough to cover the ground truth, the performance ordering between the two approaches is reversed for nonlinear problems in a strong sense. Simple ``estimate then optimize" outperforms the integrated approach in terms of stochastic dominance of the asymptotic optimality gap, i,e, the mean, all other moments, and the entire asymptotic distribution of the optimality gap is always better. Analogous results also hold under constrained settings and when contextual features are available. We also provide experimental findings to support our theory.