NetInfoF Framework: Measuring and Exploiting Network Usable Information
Authors: Meng-Chieh Lee, Haiyang Yu, Jian Zhang, Vassilis N. Ioannidis, Xiang song, Soji Adeshina, Da Zheng, Christos Faloutsos
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In our carefully designed synthetic datasets, NETINFOF correctly identifies the ground truth of NUI and is the only method being robust to all graph scenarios. Applied on real-world datasets, NETINFOF wins in 11 out of 12 times on link prediction compared to general GNN baselines. We conduct experiments by real-world graphs to answer the following research questions (RQ): RQ1. Effectiveness: How well does NETINFOF perform in real-world graphs? RQ2. Scalability: Does NETINFOF scales linearly with the input size? RQ3. Ablation Study: Are all the design choices in NETINFOF necessary? |
| Researcher Affiliation | Collaboration | Meng-Chieh Lee1, , Haiyang Yu2, Jian Zhang3, Vassilis N. Ioannidis3, Xiang Song3, Soji Adeshina3, Da Zheng3, , Christos Faloutsos1,3, 1Carnegie Mellon University, 2Texas A&M University, 3Amazon |
| Pseudocode | Yes | Algorithm 1: Compatibility Matrix with Negative Edges, Algorithm 2: NETINFOF PROBE for Link Prediction, Algorithm 3: NETINFOF PROBE for Node Classification |
| Open Source Code | Yes | Reproducibility: Code is at https://github.com/amazon-science/Network-Usable-Info-Framework. |
| Open Datasets | Yes | We evaluate NETINFOF on 7 homophily and 5 heterophily real-world graphs. ... We also conduct experiments on 3 link prediction datasets from Open Graph Benchmark (OGB) (Hu et al., 2020), namely ogbl-ddi1, ogbl-collab2, and ogbl-ppa3. ... Homophily Graphs. Cora (Motl & Schulte, 2015)4, Cite Seer (Rossi & Ahmed, 2015)5 , and Pub Med (Courtesy of the US National Library of Medicine, 1996)6 are citation networks between research articles. Computers and Photo (Ni et al., 2019)7 are Amazon co-purchasing networks between products. ogbn-ar Xiv and ogbn-Products are large graphs from OGB (Hu et al., 2020). |
| Dataset Splits | Yes | In link prediction, edges are split into training, validation and testing sets by the ratio 70%/10%/20% five times and report the average for fair comparison. In node classification, the nodes are split into training, validation and testing sets with the 2.5%/2.5%/95% ratio. |
| Hardware Specification | Yes | The experiments are conducted on an AWS EC2 G4dn instance with 192GB RAM. |
| Software Dependencies | No | The paper mentions optimizers like L-BFGS and ADAM, but does not provide specific version numbers for any software libraries or dependencies used in the experiments. |
| Experiment Setup | Yes | For small graphs, the linear models are trained by L-BFGS for 100 epochs with the patience of 5, and the non-linear models are trained by ADAM for 1000 epochs with the patience of 200. For large graphs (Products, Twitch, and Pokec), most models are trained by ADAM for 100 epochs with the patience of 10, except GPR-GNN and GAT, they are trained by ADAM for 20 epochs with the patience of 5 to speedup. All the training are full-batch, and the same amount of negative edges are randomly sampled for each batch while training. Table 14: Search space of hyperparameters. |