Neural Compositional Rule Learning for Knowledge Graph Reasoning

Authors: Kewei Cheng, Nesreen Ahmed, Yizhou Sun

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that NCRL learns high-quality rules, as well as being generalizable. Specifically, we show that NCRL is scalable, efficient, and yields state-of-the-art results for knowledge graph completion on large-scale KGs.
Researcher Affiliation Collaboration Kewei Cheng Department of Computer Science, UCLA viviancheng@cs.ucla.edu Nesreen K. Ahmed Intel Labs nesreen.k.ahmed@intel.com Yizhou Sun Department of Computer Science, UCLA yzsun@cs.ucla.edu
Pseudocode Yes Algorithm 1: Learning Algorithm
Open Source Code Yes Source code is available at https://github.com/vivian1993/NCRL.git.
Open Datasets Yes We use six widely used benchmark datasets to evaluate our NCRL in comparison to SOTA methods from knowledge graph embedding and rule learning methods. Specifically, we use the Family (Hinton et al., 1986), UMLS (Kok & Domingos, 2007), Kinship (Kok & Domingos, 2007), WN18RR (Dettmers et al., 2018), FB15K-237 (Toutanova & Chen, 2015), YAGO3-10 (Suchanek et al., 2007) datasets.
Dataset Splits No The paper mentions training and testing on different 'hops' for systematic generalization and evaluating on 'test triples', but it does not specify explicit percentages for train/validation/test splits, nor does it explicitly mention a 'validation set' or its size/percentage for the KG completion task.
Hardware Specification No The paper does not provide specific hardware details such as GPU models, CPU types, or memory specifications used for running the experiments. It only vaguely mentions 'memory capacity of our machines'.
Software Dependencies No The paper discusses models like RNN, LSTM, GRU, and Transformer, but it does not provide specific version numbers for any software libraries, frameworks (e.g., PyTorch, TensorFlow), or programming languages used.
Experiment Setup Yes In our implementation, we vary the size of the sliding window among {2, 3}. We vary the dimension of relation embeddings among {10, 100, 200, 500, 1000, 2000}. We generate k rules with the highest qualities per query relation and use them to predict missing links. We vary k among {10, 20, 40, 60, 80, 100}.