Convergence and Recovery Guarantees of the K-Subspaces Method for Subspace Clustering

Authors: Peng Wang, Huikang Liu, Anthony Man-Cho So, Laura Balzano

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

Reproducibility Variable Result LLM Response
Research Type Experimental We first conduct 3 sets of numerical tests, which correspond to K {3, 6, 9}, to examine the convergence behavior and recovery performance of the KSS method in the semi-random Uo S model (see Definition 2). ... We now conduct experiments to examine the computational efficiency and recovery accuracy of the KSS method on real datasets.
Researcher Affiliation Academia 1Department of Electrical Engineering and Computer Science, University of Michigan, Ann Arbor 2Research Institute for Interdisciplinary Sciences, School of Information Management and Engineering, Shanghai University of Finance and Economics, Shanghai 3Department of Systems Engineering and Engineering Management, The Chinese University of Hong Kong, Hong Kong.
Pseudocode Yes Algorithm 1 The TIPS initialized KSS method
Open Source Code Yes Our code is available at https://github.com/peng8wang/ ICML2022-K-Subspaces.
Open Datasets Yes We use the real datasets COIL100 (S. A. Nene & Murase, 1996a), the cropped extended Yale B (Georghiades et al., 2001), USPS (Hull, 1994), and MNIST (Le Cun, 1998). The datasets COIL-100, the cropped extended Yale B, and USPS are downloaded from http://www.cad.zju.edu. cn/home/dengcai/Data/data.html. The dataset MNIST is downloaded from https://www.csie.ntu.edu.tw/ cjlin/libsvmtools/datasets/.
Dataset Splits No No explicit mention of training, validation, or test dataset splits or cross-validation methodology. The abstract uses 'validation' in a theoretical sense, not for data partitioning.
Hardware Specification No All of our experiments are implemented in MATLAB R2020a on the Great Lakes HPC Cluster of the University of Michigan with 180GB memory and 16 cores.
Software Dependencies Yes All of our experiments are implemented in MATLAB R2020a on the Great Lakes HPC Cluster of the University of Michigan with 180GB memory and 16 cores.
Experiment Setup Yes Table 4. Parameters setting of the tested methods in the experiments . COIL20 COIL100 Yale B USPS MNIST KSS (d, τ) = (10, 0.98) (d, τ) = (10, 0.98) (d, τ) = (8, 0.98) (d, τ) = (9, 0.99) (d, τ) = (18, 0.98)