Anytime Bottom-Up Rule Learning for Knowledge Graph Completion
Authors: Christian Meilicke, Melisachew Wudage Chekol, Daniel Ruffinelli, Heiner Stuckenschmidt
IJCAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We present results for five different datasets. Three of them have been proposed as hard cases for simple (rule-based) approaches that leverage symmetries and other redundancies. Our approach performs as good as and sometimes better than most models that have been proposed recently. The results of our approach are still very good if we stop the algorithm already after a short time period. Moreover, the required resources in terms of memory and runtime are significantly smaller compared to the resources required by latent approaches. |
| Researcher Affiliation | Academia | Christian Meilicke , Melisachew Wudage Chekol , Daniel Ruffinelli and Heiner Stuckenschmidt University Mannheim {christian, mel, daniel, heiner}@informatik.uni-mannheim.de |
| Pseudocode | Yes | Algorithm 1 Anytime Bottom-up Rule Learning Any BURL(G, s, sat, Q, ts) |
| Open Source Code | Yes | The source code and datasets, used in the experiments, can be found at http://web. informatik.uni-mannheim.de/Any BURL/. |
| Open Datasets | Yes | We used in our experiments the FB15(k) dataset, its modified variant FB15-237, WN18, its modified variant WN18RR, and YAGO03-10 (in short YAGO). |
| Dataset Splits | No | We have not made use of the validation sets to find dataset specific parameter settings. |
| Hardware Specification | Yes | We conducted all our experiments on a virtual machine with 4 cores (each 2400 MHz) and 16 GB RAM. |
| Software Dependencies | No | Any BURL is written in Java and requires no external libraries. |
| Experiment Setup | Yes | For all our experiments we used exactly the same parameter setting. We have chosen the quality criteria Q to allow only those rules that generate at least two correct predictions, which is a very lax criteria. We have set the required saturation rate sat to 99%. We set ts = 1 second, pc = 5, and s = 500, which is the sample size that determines the precision of the confidence value. As default setting we used the maximum aggregation strategy. |