Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Subsampled Ensemble Can Improve Generalization Tail Exponentially
Authors: Huajie Qian, Donghao Ying, Henry Lam, Wotao Yin
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The major claim made in our paper is that we proposed a new ensemble learning method that attains an exponentially decaying tail for excess risk. This claim is theoretically proved in Section 2. Moreover, we have conducted extensive numerical experiments in Section 3 to support our theoretical results. |
| Researcher Affiliation | Collaboration | Huajie Qian DAMO Academy Alibaba Group Bellevue, WA 98004 EMAIL Donghao Ying IEOR Department UC Berkeley Berkeley, CA 94720 EMAIL Henry Lam IEOR Department Columbia University New York, NY 10027 EMAIL Wotao Yin DAMO Academy Alibaba Group Bellevue, WA 98004 EMAIL |
| Pseudocode | Yes | Algorithm 1 Majority Vote Ensembling (Mo VE) Algorithm 2 Retrieval and ϵ-Optimality Vote Ensembling (ROVE / ROVEs) Algorithm 3 Majority Vote Ensembling for Set-Valued Learning Algorithms |
| Open Source Code | Yes | The code is available at: https://github.com/mickeyhqian/Vote Ensemble. |
| Open Datasets | Yes | We use three datasets from the UCI Machine Learning Repository [7]: Bike Sharing [21], Superconductivity [33], and Gas Turbine Emission [1]. |
| Dataset Splits | Yes | For neural networks, the base learner splits the data into training (70%) and validation (30%) To evaluate the tail probabilities of out-of-sample costs, we permute each dataset 100 times, and each time use the first half for training and the second for testing. |
| Hardware Specification | No | All experiments are conducted on a personal computer, and Gurobi Optimizer is required for certain experiments on stochastic programs. The code is available at: https://github.com/mickeyhqian/Vote Ensemble. |
| Software Dependencies | No | All experiments are conducted on a personal computer, and Gurobi Optimizer is required for certain experiments on stochastic programs. |
| Experiment Setup | Yes | For neural networks, the base learner splits the data into training (70%) and validation (30%), and uses Adam to minimize the mean squared error (MSE), with early stopping triggered when the validation improvement falls below 3% between epochs. Recommended Configurations... For discrete space Θ, use k = max(10, n/200), B = 200 for Mo VE, and k1 = k2 = max(10, n/200), B1 = 20, B2 = 200 for ROVE and ROVEs. For continuous space Θ, use k1 = max(30, n/2), k2 = max(30, n/200), B1 = 50, B2 = 200 for ROVE and ROVEs. The ϵ in ROVE and ROVEs is selected such that maxθ S(1/B2) PB2 b=1 1(θ bΘϵ,b k2 ) 1/2. |