A Model-Agnostic Randomized Learning Framework based on Random Hypothesis Subspace Sampling

Authors: Yiting Cao, Chao Lan

ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show they converge efficiently as k increases and outperform their model-specific counterparts including random fourier feature, random vector functional link and extra tree on real-world data sets.
Researcher Affiliation Academia School of Computer Science, University of Oklahoma, Norman, OK, USA. Correspondence to: Chao Lan <clan@ou.edu>.
Pseudocode Yes Framework 1 The RHSS-based Learning Framework
Open Source Code Yes The codes of all experiments are available at https://github.com/yxc827/RHSS.
Open Datasets Yes We compare the performance of the proposed RHSS-KRR, RHSS-MLP and RHSS-Tree with their existing randomized counterparts on three public real-world data sets, namely, Crime and Community, Adult and COMPAS.
Dataset Splits No On each data set, we use the first half of the instances for training and the other half for testing.
Hardware Specification No The paper does not explicitly mention specific hardware details such as GPU/CPU models, processors, or memory used for experiments.
Software Dependencies No The paper does not provide specific software dependency versions (e.g., library names with version numbers) needed to replicate the experiment.
Experiment Setup Yes For RHSS-KRR, the hypothesis coefficients are sampled independently from N(0, 1). ... the regularization coefficient is optimized to 0.1 on Crime and 0.001 on the other two data sets. ... single-hidden-layer architecture with 20 hidden neurons and ReLU activation function. ... regularization coefficient is optimized to 10. ... all non-optimized parameters are independently sampled from N(0, 1). ... bootstrap sample size is set to 80% of the original training set. ... k is the projected dimension of RP, and we fix the number of sampled hypotheses to 100 for RHSS-KRR. For RP, all projection entries are independently sampled from N(0, 0.01).