Large-Scale Graph-Based Semi-Supervised Learning via Tree Laplacian Solver
Authors: Yan-Ming Zhang, Xu-Yao Zhang, Xiao-Tong Yuan, Cheng-Lin Liu
AAAI 2016 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments show the significant scalability improvement over existing scalable semi-supervised learning methods. Here we conduct experiments to evaluate our method in accuracy and speed. |
| Researcher Affiliation | Academia | Yan-Ming Zhang and Xu-Yao Zhang National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China Xiao-Tong Yuan Jiangsu Province Key Laboratory of Big Data Analysis Technology, Nanjing University of Information Science and Technology, Nanjing, China Cheng-Lin Liu National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences, Beijing, China Center of Excellence for Brain Science and Intelligence Technology, Chinese Academy of Sciences, Beijing, China |
| Pseudocode | Yes | Algorithm 1 Tree Laplacian Factorization. Algorithm 2 Linear Equations System Solver. |
| Open Source Code | Yes | The codes of Tb TL are available at www.nlpr.ia.ac.cn/pal/ymzhang/index.html . |
| Open Datasets | Yes | We use 4 image data sets (COIL20, USPS, MNIST and Letter) and 2 text data sets (RCV1 and News20) to test these methods. Their basic properties are listed in Table 1. |
| Dataset Splits | Yes | For each data set, we randomly select l data points as labeled data and run different methods to predict the unlabeled data. To control the variance, we repeat the procedure 20 times for different labeled/unlabeled splits. We let l vary from n {1%, 2%, 4%, 8%, 16%, 32%, 64%}, and report the results in Figure 1. |
| Hardware Specification | Yes | All methods are run on a PC with a 3.10 GHz 4-core CPU and 8 GB RAM. |
| Software Dependencies | No | The paper mentions using specific algorithms like CMG and methods like conjugate gradient (CG) but does not provide version numbers for any software dependencies. |
| Experiment Setup | Yes | For Tb TL, we define C as Cii = 100 for 1 i l, Cii = 0 for l +1 i n. MST is used for both GPA and Tb TL. For Eigen, we approximately compute the 400 smallest eigenvectors. For AGR, 1000 prototypes are selected by k-means clustering, and s is chosen from {2, 4, 8, 16} by minimizing the test error. |