Solving Sparse \& High-Dimensional-Output Regression via Compression

Authors: Renyuan Li, Zhehui Chen, Guanyi Wang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirically, numerical results further validate the theoretical findings, showcasing the efficiency and accuracy of the proposed framework.
Researcher Affiliation Collaboration Renyuan Li Department of Industrial Systems Engineering & Management National University of Singapore renyuan.li@u.nus.edu Zhehui Chen Google zhehuichen@google.com Guanyi Wang Department of Industrial Systems Engineering & Management National University of Singapore guanyi.w@nus.edu.sg
Pseudocode Yes Algorithm 1 Projected Gradient Descent (for Second Stage) (...) Algorithm 2 Implemented Projected Gradient Descent (for Second Stage)
Open Source Code Yes The implemented code could be found on Github https://github.com/from-ryan/Solving_ SHORE_via_compression.
Open Datasets Yes We select two benchmark datasets in multi-label classification, Wiki10-31K and EURLex-4K[5] due to their sparsity property.
Dataset Splits No The paper mentions splitting the synthetic data into a training set and a testing set ('training set Stra with 80% and a testing set Stest with rest 20%'), but does not explicitly mention a separate validation set for hyperparameter tuning or model selection.
Hardware Specification Yes All experiments are conducted in Dell workstation Precision 7920 with a 3GHz 48Cores Intel Xeon CPU and 128GB 2934MHz DDR4 Memory.
Software Dependencies Yes The proposed method and other methods are solved using Py Torch version 2.3.0 and scikit-learn version 1.4.2 in Python 3.12.3.
Experiment Setup Yes Parameter setting. For synthetic data, we set input dimension d = 104, output dimension K = 2 104, and sparsity-level s = 3. We generate in total n = 3 104, i.i.d. samples (...) We select the number of rows for compressed matrix Φ by m {100, 300, 500, 700, 1000, 2000}. (...) For evaluating the proposed prediction method, Algorithm 2, we pick a fixed stepsize η = 0.9, F = RK + , and set the maximum iteration number as T = 60, and run prediction methods over the set Stest.