Scalable Rule Learning via Learning Representation
Authors: Pouya Ghiasnezhad Omran, Kewen Wang, Zhe Wang
IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our system outperforms some state-of-the-art systems. Specifically, for massive KGs with hundreds of predicates and over 10M facts, RLv LR is much faster and can learn much more quality rules than major systems for rule learning in KGs such as AMIE+. We also used the RLv LR-mined rules in an inference module to carry out the link prediction task. In this task, RLv LR outperformed Neural LP, a state-of-the-art link prediction system, in both runtime and accuracy. |
| Researcher Affiliation | Academia | Pouya Ghiasnezhad Omran, Kewen Wang, Zhe Wang Griffith University pouya.ghiasnezhadomran@griffithuni.edu.au, {k.wang, zhe.wang}@griffith.edu.au |
| Pseudocode | Yes | Algorithm 1 Learn rules for a KG and a target predicate Input: a KG K, a predicate Pt, an integer len 2, and two real numbers Min SC, Min HC [0, 1] Output: a set Rule of CP rules 1: K := Sampling(K, Pt, len) 2: (P, A) := Embeddings(K ) 3: Candidates := 4: for 2 l len do 5: Add Rule Search(K , Pt, P, A, l) to Candidates 6: end for 7: Rules := Evaluate(Candidates, K) 8: Rules := Filter(Candidates, Min SC, Min HC) 9: return Rules |
| Open Source Code | Yes | The executable codes, benchmark datasets and experimental results are publicly available at https://www.ict.griffith.edu.au/aist/RLv LR/. |
| Open Datasets | Yes | The executable codes, benchmark datasets and experimental results are publicly available at https://www.ict.griffith.edu.au/aist/RLv LR/. The benchmark datasets adopted in our experiments include various versions of Freebase, YAGO, DBpedia and Wikidata that are widely used in experimental evaluations by major systems for rule learning and link prediction in KGs. FB15K-237 [Toutanova and Chen, 2015] (aka. FB15KSelected) and FB75K (from NIPS 13 dataset) are obtained from Freebase and widely adopted for link prediction benchmarking [Yang et al., 2017]. |
| Dataset Splits | No | Each dataset is divided into training set (70%) and test set (30%). |
| Hardware Specification | Yes | For benchmark KGs FB15K-237 and FB75K, we used a PC with Intel Core i5-4590 CPU at 3.30GHz 4 and with 5GB of RAM, running Ubuntu 14.04. For other larger benchmark KGs we tested, the experiments were conducted on a server with Intel Xeon CPU at 2.67GHz (one thread) and with 40GB of RAM, running Red Hat Linux 6.1. |
| Software Dependencies | No | The paper mentions operating systems (Ubuntu 14.04, Red Hat Linux 6.1) but does not list specific software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions). |
| Experiment Setup | Yes | The rule quality was measured by standard confidence (SC) and head coverage (HC) [Gal arraga et al., 2015]. Experiment 1: #R, SC 0.1 and HC 0.01. #QR, SC 0.7. Experiment 2: We set a 5 hours time limit, SC 0.01 and HC 0.001. Section 6.2: The parameters of RLv LR were set to SC 0.005 and HC 0.001 to achieve better accuracy. |