A*Net: A Scalable Path-based Reasoning Approach for Knowledge Graphs

Authors: Zhaocheng Zhu, Xinyu Yuan, Michael Galkin, Louis-Pascal Xhonneux, Ming Zhang, Maxime Gazeau, Jian Tang

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on both transductive and inductive knowledge graph reasoning benchmarks show that A*Net achieves competitive performance with existing state-of-the-art path-based methods, while merely visiting 10% nodes and 10% edges at each iteration. On a million-scale dataset ogbl-wikikg2, A*Net not only achieves a new state-of-the-art result, but also converges faster than embedding methods.
Researcher Affiliation Collaboration 1Mila Québec AI Institute, 2University of Montréal 3Intel AI Lab, 4Peking University, 5LG Electronics AI Lab 6HEC Montréal, 7CIFAR AI Chair
Pseudocode Yes Algorithm 1 A*Net
Open Source Code Yes Code is available at https://github.com/DeepGraphLearning/AStarNet
Open Datasets Yes We evaluate A*Net on 4 standard knowledge graphs, FB15k-237 [40], WN18RR [16], YAGO3-10 [30] and ogbl-wikikg2 [25].
Dataset Splits Yes For the transductive setting, we use the standard splits from their original works [40, 16]. For the inductive setting, we use the splits provided by [39], which contains 4 different versions for each dataset.
Hardware Specification Yes We train A*Net with 4 Tesla A100 GPUs (40 GB)
Software Dependencies No The paper mentions developing based on an open-source codebase but does not specify version numbers for any software dependencies like Python, PyTorch, or DGL.
Experiment Setup Yes For the neural priority function, we have two hyperparameters: K for the maximum number of nodes and L for the maximum number of edges. To make hyperparameter tuning easier, we define maximum node ratio α = K/|V| and maximum average degree ratio β = L|V|/K|E|, and tune the ratios for each dataset. The maximum edge ratio is determined by αβ. The other hyperparameters are kept the same as the values in [58].