Node2ket: Efficient High-Dimensional Network Embedding in Quantum Hilbert Space
Authors: Hao Xiong, Yehui Tang, Yunlin He, Wei Tan, Junchi Yan
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the experiments, we show diverse virtues of our methods, including but not limited to: the overwhelming performance on downstream tasks against conventional low-dimensional NE baselines with the similar amount of computing resources |
| Researcher Affiliation | Academia | Hao Xiong, Yehui Tang, Yunlin He, Wei Tan, Junchi Yan Department of Computer Science and Engineering, Shanghai Jiao Tong University |
| Pseudocode | Yes | Algorithm 1: An iteration of node2ket. |
| Open Source Code | Yes | The used data, the source code of node2ket, node2ket+, and the LIBN2K library, the compilation instructions, and all the scripts to run experiments in Sec. 6, are provided in the link shown right below author names in the first page. https://github.com/Shaw Xh/node2ket |
| Open Datasets | Yes | We study five public real-world datasets of different scales, different densities, and containing different types of information. All of the networks are processed as undirected ones. The detailed statistics of the networks are given in Table 5. We give a brief introduction of the datasets as follows: You Tube (Tang & Liu, 2009b) is a social network... Blog Catalog (Tang & Liu, 2009a) (BC) is a social network... PPI (Breitkreutz et al., 2007) is a subgraph... Arxiv GR-QC (Leskovec et al., 2007) is a collaboration network... DBLP (Ley, 2002) is a citation network... |
| Dataset Splits | Yes | For evaluation, we divide all the labels into train and test sets, and then we train an independent classifier to classify the nodes according to nodes embeddings and labels in the training set. Then the trained classifier is used to classify the nodes in the test set by the learned node embeddings. Concretely, we conduct experiments on PPI and You Tube. After embeddings are obtained, LIBLINEAR (Fan et al., 2008) is used to train one-vs-rest Logistic Regression classifiers. We range the portion of training data from 1% to 9% for You Tube and 10% to 90% for PPI. |
| Hardware Specification | Yes | All of the experiments are run on a single machine with Intel(R) Core(TM) i9-7920X CPU @ 2.90GHz with 24 logical cores and 128GB memory. |
| Software Dependencies | No | After embeddings are obtained, LIBLINEAR (Fan et al., 2008) is used to train one-vs-rest Logistic Regression classifiers. The version number for LIBLINEAR is not explicitly provided. |
| Experiment Setup | Yes | For all methods that can set the number of threads... we use 8 threads as the default. For our method, we use the MT objective on the tasks of NR and LP, and the SGNS objective on the NC task. For the MT objective, the margin is set as γ = 0.1, and for the SGNS objective, the number of negative samples is set as K = 5. We use random walk with window size w = 2 for LP, w = 1 for NR and NC. The default Riemannian optimization order is 0. The default sampling number is 100M. For You Tube the sampling number is 1000M. Default input is data in the network form. Default embedding type is type-I. |