LogicDP: Creating Labels for Graph Data via Inductive Logic Programming

Authors: Yuan Yang, Faramarz Fekri, James Clayton Kerce, Ali Payani

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

Reproducibility Variable Result LLM Response
Research Type Experimental In the experiments, we demonstrate the following properties of LOGICDP: (P1) the budget-aware framework empowers LOGICDP to generate high-quality triples leading to better data efficiency than generic weakly-supervised methods; (P2) LOGICDP is robust against the incorrect labels; and (P3) the logic rule is interpretable, allowing humans to supervise the model efficiently.
Researcher Affiliation Collaboration Yuan Yang1, Faramarz Fekri1, James C. Kerce1 & Ali Payani2 Georgia Institute of Technology1, Cisco2 {yyang754@,faramarz.fekri@ece.,clayton.kerce@gtri.}gatech.edu apayani@cisco.com
Pseudocode Yes Algorithm 1: Training graph reasoning model with LOGICDP
Open Source Code Yes Implementation available here 1. https://github.com/gblackout/logic-data-programming
Open Datasets Yes We evaluate LOGICDP on two KG datasets. FB15K-237 is a subset of the Freebase dataset (Toutanova & Chen, 2015) which contains general knowledge facts. WN18RR (Dettmers et al., 2018) is the subset of Word Net18 which contains relations between English words. We also evaluate LOGICDP on the Visual Genome (VG) dataset provided in (Krishna et al., 2016)
Dataset Splits Yes For FB15K and WN18RR, we evaluate the LOGICDP on standard graph reasoning on binary predicates with train/valid/test splits provided in (Yang et al., 2017). The bootstrap set Tinit for each of the 70 classes contains 15 random triples, and the valid, test, and unlabeled sets are split with a 10%/80%/10% ratio on the remaining triples.
Hardware Specification Yes Experiments are conducted on a machine with i7-8700K and one GTX1080ti.
Software Dependencies No The paper does not provide specific software dependencies with version numbers, such as Python or PyTorch versions.
Experiment Setup Yes We set hyper-parameters γ = 0.7, and C = 20 by running a grid search over the validation set. For FB15K and WN18RR, we evaluate the mean reciprocal rank (MRR) and hits@10. For the scene graph dataset, We evaluate the Recall@1 (R@1) score.