Cooperation of Experts: Fusing Heterogeneous Information with Large Margin
Authors: Shuo Wang, Shunyang Huang, Jinghui Yuan, Zhixiang Shen, Zhao Kang
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Rigorous theoretical analyses guarantee the framework s feasibility and stability, while extensive experiments across diverse benchmarks demonstrate its superior performance and broad applicability. Our code is available at https: //github.com/strange Alan/Co E. |
| Researcher Affiliation | Academia | 1University of Electronic Science and Technology of China, Chengdu, Sichuan Province, China 2Northwestern Polytechnical University. Correspondence to: Shuo Wang <EMAIL>, Zhao Kang <EMAIL>. |
| Pseudocode | No | The paper describes its methods using mathematical equations and textual explanations within Section 4 'Our Proposed Method' and Section 5 'Theoretical Analysis'. It does not contain any explicit pseudocode or algorithm blocks with numbered steps formatted as code. |
| Open Source Code | Yes | Our code is available at https: //github.com/strange Alan/Co E. |
| Open Datasets | Yes | To thoroughly evaluate the effectiveness of Co E, we conduct experiments on five benchmark network datasets, including citation networks (ACM (Yun et al., 2019) and DBLP (Yun et al., 2019)), review networks (Yelp (Lu et al., 2019) and Amazon (Mc Auley & Leskovec, 2013; Gao et al., 2023)), and a large-scale citation network MAG (Wang et al., 2020). Additionally, we perform experiments on four multimodal datasets: ESP, Flickr, IAPR, and NUS (Xia et al., 2023), which lack structural information. |
| Dataset Splits | Yes | Table 6. Statistics of multi-relational network datasets. Dataset Nodes Relation type Edges Features Classes Training Validation Test ACM 3,025 Paper-Author-Paper (PAP) 26,416 1,902 3 600 300 2,125 ... Table 7. Statistics of multimodal datasets. Dataset Classes Total Features Modal Training Validation Test IAPR 6 7,855 100 Image,text 3,926 1,961 1,968 |
| Hardware Specification | Yes | All experiments are conducted on a platform equipped with an Intel(R) Xeon(R) Gold 5220 CPU and an NVIDIA A800 80GB GPU, using Py Torch 2.1.1 and DGL 2.4.0. |
| Software Dependencies | Yes | All experiments are conducted on a platform equipped with an Intel(R) Xeon(R) Gold 5220 CPU and an NVIDIA A800 80GB GPU, using Py Torch 2.1.1 and DGL 2.4.0. |
| Experiment Setup | Yes | Our model is trained using the Adam optimizer. The hyper-parameter settings for all datasets are presented in Table 5. Here, E represents the number of training epochs, which is tuned within the set {100, 200, 400, 500, 800, 1000} and lr is the learning rate which is searched within the set {0.0001, 0.005, 0.001, 0.005, 0.01}. The hidden layer dimension dh and the representation dimension d of the graph encoder GCN are tuned within the set {32, 64, 128, 256}. The number of neighbors K for KNN is searched within the set {5, 10, 15, 20, 30}. The order of graph aggregation r and the number of layers L in GCN are set to either 2 or 3, which is in line with the common layer count of GNN models (Baranwal et al.). The temperature parameter τc in the contrastive loss is fixed at 0.2. In the optimization phase, the parameters α and γ are both set to 100. |