KMF: Knowledge-Aware Multi-Faceted Representation Learning for Zero-Shot Node Classification
Authors: Likang Wu, Junji Jiang, Hongke Zhao, Hao Wang, Defu Lian, Mengdi Zhang, Enhong Chen
IJCAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Finally, we conduct extensive experiments on several public graph datasets and design an application of zero-shot cross-domain recommendation. The quantitative results demonstrate both the effectiveness and generalization of KMF with the comparison of state-of-the-art baselines. |
| Researcher Affiliation | Collaboration | 1University of Science and Technology of China , 2State Key Laboratory of Cognitive Intelligence, China , 3Fudan University , 4Tianjin University , 5Meituan-Dianping Group wulk@mail.ustc.edu.cn, jjjiang22@m.fudan.edu.cn, hongke@tju.edu.cn, {wanghao3,liandefu,cheneh}@ustc.edu.cn, zhangmengdi02@meituan.com |
| Pseudocode | No | The paper does not contain any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not provide a statement or link indicating that the source code for the methodology is openly available. |
| Open Datasets | Yes | We conduct experiments on three public real-world scholar datasets, DBLP, M10 [Pan et al., 2016], and Cora E (Cora Enrich) [Bojchevski and G unnemann, ]. |
| Dataset Splits | Yes | Class Split II: the seen classes are partitioned into train and validation parts, and the unseen classes are still used for testing. The [ train/val/test ] class split for Cora E, M10, and DBLP are [2/2/3], [3/3/4], and [2/2/2]. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used for running its experiments. |
| Software Dependencies | No | The paper mentions tools like 'hyperopt' and 'BERT' but does not provide specific version numbers for software dependencies. |
| Experiment Setup | Yes | The embedding size of ei produced by BERT is 1,024. The radius of topic neighbors R = {0, 1, 2, 3}. We set the layer L = [0, 1, ..., 4] of GCN for compositional embedding, the dimension of hidden state and output db = 1024, the maximum of N(v) is set to 10, and the sampling number Q = [0, 5, 10, 15, 20] of negative contrastive pairs in graph contrastive learning. The magnitude parameters pm, pτ uniform[0.1, 0.5] in the topic mask sampling. In the geometric constraints, dr, rτ uniform[0.1, 0.5] as well. For other hyperparameters, we set the attenuation coefficient α = 0.8, temperature τ = 10, filtering soft cut %P = 25%, and tradeoff λ1,2 uniform[0.1, 0.5] according to the feedback of experimental performance. We adopt Adam [Kingma and Ba, 2015] with a learning rate of 0.001 to optimize our model. |