Partial-Label and Structure-constrained Deep Coupled Factorization Network
Authors: Yan Zhang, Zhao Zhang, Yang Wang, Zheng Zhang, Li Zhang, Shuicheng Yan, Meng Wang10948-10955
AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments on image datasets demonstrate the effectiveness of DS2CF-Net for representation learning and clustering. |
| Researcher Affiliation | Collaboration | Yan Zhang,1 Zhao Zhang, 1,2 Yang Wang, 2 Zheng Zhang,3 Li Zhang, 1 Shuicheng Yan, 4 Meng Wang 2 1 School of Computer Science and Technology, Soochow University, Suzhou 215006, China 2 School of Computer Science and Information Engineering, Hefei University of Technology, Hefei, China 3 Harbin Institute of Technology & Peng Cheng Laboratory, Shenzhen, China 4 YITU Technology |
| Pseudocode | Yes | Algorithm 1. Optimization procedures of DS2CF-Net Inputs: Partially labeled data matrix X = [XL, XU], the constant r and tunable parameters α, β and γ. Initialization: t = 0; Initialize W and Z to be random matrices; Initialize P and A by labeled data; Initialize QU by the cosine similarities over XU; Initialize SU by the cosine similarities over X; Initialize SV using semi-supervised weights, that is, supervised ones for XL and cosine similarities for XU. For each fixed number m of layers: While not converged do 1. Update W t+1 m and Zt+1 m by Eqs.(14-15), and then we can obtain V t+1 m = AZ0 . . . Zt+1 m ; 2. Update (SU)t+1 and (SV )t+1 by Eqs.(16-17); 3. Update P t+1 by Eq.(18), update the estimated soft labels of XU as XT U P t+1, and then update AU by Eq.(9); 4. Update the label constraint matrix A by Eq.(5); 5. Update QU using cosine similarities based on (V t+1 m )i i {l + 1, . . . , N}, and update matrix Q; 6. Check the convergence conditions: if W t+1 m W t m 2 F E and V t+1 m V t m 2 F E, stop; else t = t + 1. End While End for Outputs: Deep low-dimensional representation V m. |
| Open Source Code | No | The paper does not provide any explicit statement or link indicating that the source code for the described methodology is publicly available. |
| Open Datasets | Yes | In this study, 4 public databases are involved, i.e., AR (Bergstra et al. 2013), ETH80 (Leibe et al. 2003), USPS (Hull 1994) and Fashion MNIST (Xiao et al. 2017). Detailed information of the used databases is described in Table 1. |
| Dataset Splits | Yes | For fair comparison, we randomly choose 40% labeled samples per class for semi-supervised models and set the number of layers to 3 for deep models. and For each database, the labeled proportion varies from 10% to 90% and we randomly choose 3 categories. |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used for running the experiments. |
| Software Dependencies | No | The paper does not provide specific software dependencies with version numbers required to replicate the experiment. |
| Experiment Setup | Yes | The rank of the representation is set to K+1 for clustering as (Liu et al. 2014; Zhang et al.2019a). The number of layers is set to 3 for deep models. Specifically, we first fix γ = 1 to tune α and β from {10-5, 10-4,..., 105}. Then, we use selected α and β to tune γ. |