Differentiable Unsupervised Feature Selection based on a Gated Laplacian

Authors: Ofir Lindenbaum, Uri Shaham, Erez Peterfreund, Jonathan Svirsky, Nicolas Casey, Yuval Kluger

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Using several real-world examples, we demonstrate the efficacy and advantage of the proposed approach over leading baselines.
Researcher Affiliation Academia Ofir Lindenbaum Faculty of Engineering Bar-Ilan University Ramat Gan, Israel 5290002 ofir.lindenbaum@biu.ac.il Uri Shaham Center for Outcome Research and Evaluation Yale University New Haven, CT 06510, USA uri.shaham@yale.edu Erez Peterfreund Hebrew University Jonathan Svirsky Independent Researcher Nicolas Casey University of Pennsylvania Yuval Kluger Program in Applied Math Department of Pathology Yale University New Haven, CT 06510, USA yuval.kluger@yale.edu
Pseudocode Yes Algorithm 1 Differentiable Unsupervised Feature Selection (DUFS) Pseudo-code
Open Source Code Yes Code is available in the supplemental material.
Open Datasets Yes All datasets are publicly available, see description in Appendix section S7
Dataset Splits No The paper describes procedures for evaluating clustering accuracy, such as averaging results over 20 runs, but it does not specify explicit training, validation, or testing dataset splits for the DUFS method or the clustering task.
Hardware Specification Yes Description of the computational resources are described in Section S3 in the Appendix.
Software Dependencies No The paper indicates computational resources are detailed in Appendix S3, but the main text does not provide specific software names along with their version numbers required for reproducibility.
Experiment Setup Yes At initialization µi = 0.5 for i = 1, ..., d. ... Our algorithm involves applying a standard optimization scheme (such as stochastic gradient decent) to objective (8) or (9). ... We used t = 2, which was observed to improve the performance of our proposed approach. ... We perform k-means clustering using the leading 50, 100, 150, 200, 250, or 300 selected features and average the results over 20 runs. ... The number clusters k is set as the number of classes. ... Description of all training procedure appears in Section S3 in the Appendix.