Few-shot Relational Reasoning via Connection Subgraph Pretraining

Authors: Qian Huang, Hongyu Ren, Jure Leskovec

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on real KGs, including NELL, FB15K-237, and Concept Net, demonstrate the effectiveness of our framework: we show that even a learning-free implementation of CSR can already perform competitively to existing methods on target few-shot tasks; with pretraining, CSR can achieve significant gains of up to 52% on the more challenging inductive few-shot tasks where the entities are also unseen during (pre)training.
Researcher Affiliation Academia Qian Huang Stanford University qhwang@cs.stanford.edu Hongyu Ren Stanford University hyren@cs.stanford.edu Jure Leskovec Stanford University jure@cs.stanford.edu
Pseudocode Yes Algorithm 1 Hypothesis Proposal Module Mp of CSR-GNN Algorithm 2 CSR-GNN Full Architecture
Open Source Code Yes The implementation of CSR can be found in https://github.com/snap-stanford/csr.
Open Datasets Yes We evaluate our method CSR-OPT and CSR-GNN on few-shot KG completion tasks over three real world KGs NELL, FB15K-237 and Concept Net. We take NELL directly from NELL-One [23]... FB15K-237 and Concept Net following [12, 23].
Dataset Splits Yes We summarize the statistics of all three datasets in Table 1. BG refers to the background KG available during training time. Trans-BG, Ind-BG, Ind-Test.
Hardware Specification No The paper states: "Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)? [Yes] See Appendix." However, no specific hardware details (like GPU/CPU models, memory, or specific cloud instance types) are provided in the main body of the paper.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python 3.x, PyTorch 1.x) in its main text.
Experiment Setup Yes For simplicity, we only consider the number of few shot example K = 3, even though all methods here can generalize to arbitrary K. See appendix A for more details on full experiment setups.