Long-range Meta-path Search on Large-scale Heterogeneous Graphs

Authors: Chao Li, Zijie Guo, qiuting he, Kun He

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments across diverse heterogeneous datasets validate LMSPS s capability in discovering effective long-range meta-paths, surpassing state-of-the-art methods.
Researcher Affiliation Collaboration 1School of Computer Science and Technology, Huazhong University of Science and Technology 2 China Mobile Information Technology Co.,Ltd. 3 School of Computer Science, Fudan University
Pseudocode Yes Algorithm 1 The search algorithm of LMSPS
Open Source Code Yes Our code is available at https://github.com/JHL-HUST/LMSPS.
Open Datasets Yes We evaluate LMSPS on several representative heterogeneous graph datasets, including DBLP, IMDB, ACM, and Freebase from HGB benchmark [34], and the large-scale dataset OGBN-MAG from OGB challenge [20].
Dataset Splits Yes target type nodes are divided into 24% for training, 6% for validation, and 70% for testing. ... For the OGBN-MAG dataset, we use the official data partition, where papers published before 2018, in 2018, and since 2019 are nodes for training, validation, and testing, respectively.
Hardware Specification Yes We use Pytorch [38] to run all experiments on one Tesla V100 GPU with 16GB GPU memory.
Software Dependencies No The paper mentions "Pytorch [38]" but does not specify a version number for Pytorch or any other software dependencies, such as specific Python libraries or CUDA versions.
Experiment Setup Yes We set the number of selected meta-paths M = 30 for all datasets. The final search space V = 60. The maximum hop is 6 for ogbn-mag, DBLP, 5 for IMDB, ACM, and 3 for Freebase. ... For searching in the super-net, we train for 200 epochs. ... τ linearly decays with the number of epochs from 8 to 4. The learning rate is 0.001 for all search stages and HGB training stage, and 0.003 for OGBN-MAG training stage. The weight decay is always 0.