Pure Message Passing Can Estimate Common Neighbor for Link Prediction

Authors: Kaiwen Dong, Zhichun Guo, Nitesh Chawla

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments on benchmark datasets from various domains, where our method consistently outperforms the baseline methods, establishing new state-of-the-arts.Our empirical investigations provide compelling evidence of MPLP s dominance. Benchmark tests reveal that MPLP not only holds its own but outstrips state-of-the-art models in link prediction performance.
Researcher Affiliation Academia 1Computer Science and Engineering, University of Notre Dame 2Lucy Family Institute for Data and Society, University of Notre Dame
Pseudocode No The paper describes methods using mathematical equations and textual explanations (e.g., Equation 3, Equation 5), but does not include a formally labeled 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes Our code is publicly available at https://github.com/Barcavin/efficient-node-labelling.
Open Datasets Yes We conduct evaluations across a diverse spectrum of 15 graph benchmark datasets, which include 8 non-attributed and 7 attributed graphs. It also includes three datasets from OGB [10] with predefined train/test splits. ... USAir [44]: a graph of US airlines; NS [45]: a collaboration network of network science researchers; PB [46]: a graph of links between web pages on US political topics; Yeast [47]: a protein-protein interaction network in yeast;
Dataset Splits Yes In the absence of predefined splits, links are partitioned into train, validation, and test sets using a 70-10-20 percent split.
Hardware Specification Yes We run our experiments on a Linux system equipped with an NVIDIA A100 GPU with 80GB of memory.
Software Dependencies No The paper states 'We implement MPLP in Pytorch Geometric framework [52]' but does not provide specific version numbers for Pytorch Geometric or other software dependencies.
Experiment Setup Yes The chosen hyperparameters are as follows: Number of Hops (r): We set the maximum number of hops to r = 2. Node Signature Dimension (F): The dimension of node signatures, F, is fixed at 1024, except for Citation2 with 512. The minimum degree of nodes to be considered as hubs (b): We experiment with values in the set [50, 100, 150]. Batch Size (B): We vary the batch size depending on the graph type: For the 8 nonattributed graphs, we explore batch sizes within [512, 1024]. For the 4 attributed graphs coming from [51], we search within [2048, 4096]. For OGB datasets, we use 32768 for Collab and PPA, and 261424 for Citation2.