ANRL: Attributed Network Representation Learning via Deep Neural Networks
Authors: Zhen Zhang, Hongxia Yang, Jiajun Bu, Sheng Zhou, Pinggang Yu, Jianwei Zhang, Martin Ester, Can Wang
IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on six real-world networks, including two social networks, two citation networks and two user behavior networks. The results empirically show that ANRL can achieve relatively significant gains in node classification and link prediction tasks. |
| Researcher Affiliation | Collaboration | Zhen Zhang1,2, Hongxia Yang3, Jiajun Bu1, Sheng Zhou1,2, Pinggang Yu1,2, Jianwei Zhang3, Martin Ester4, Can Wang1 1 College of Computer Science, Zhejiang University, China 2Alibaba-Zhejiang University Joint Institute of Frontier Technologies, China 3 Alibaba Group, China 4 Simon Fraser University, Canada |
| Pseudocode | Yes | Algorithm 1 Joint ANRL Learning Framework |
| Open Source Code | No | The paper does not explicitly provide a link to the open-source code for the ANRL method. It mentions the use of implementations released by original authors for baseline comparisons but not for their own proposed model. |
| Open Datasets | Yes | We summarize statistics of the six datasets in Table 1 with more descriptions as follows: Social Network. Facebook1 [Leskovec and Mcauley, 2012] and UNC [Traud et al., 2012] datasets are two typical social networks used in [Grover and Leskovec, 2016; Liao et al., 2017]. Citation Network. Citeseer and Pubmed 2 which are used in [Yang et al., 2016] consist of bibliography publication data. The edge represents that each paper may cite or be cited by other papers. |
| Dataset Splits | Yes | For a given dataset, we randomly select 20 samples from each class and treat them as the labeled data to train semi-supervised baselines following the same strategy in [Yang et al., 2016]. After having obtained the node representations, we randomly sample 30% labeled nodes to train a SVM classifier and the rest of the nodes are used to test performances. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models, memory, or types of computing resources used for experiments. |
| Software Dependencies | No | The paper mentions using implementations released by original authors for baselines but does not specify any software names with version numbers for their own model or the experimental setup (e.g., Python version, specific libraries, or frameworks). |
| Experiment Setup | Yes | We set the embedding size d as 64 in Fraud Detection dataset and 128 for the remaining datasets. For LINE, we concatenate both first-order and second-order as our final representations. Furthermore, we set window size b as 10, walk length l as 80, walks per node γ as 10, negative samples as 5. For ANRL, the number of layers and dimensions for left output branch are shown in Table 2 and we only use one layer in right output branch. |