Representation Learning Based Predicate Invention on Knowledge Graphs
Authors: Man Zhu, Pengfei Huang, Lei Gu, Xiaolong Xu, Jingyu Han
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate the proposed approach on FB15k and DRKG datasets, and the results demonstrate the effectiveness of SEm PI in discovering new predicates. Experiments We have designed and executed a series of experiments to address the following inquiries: [Q1] Is the scoring function effective in capturing both semantics in embeddings and regularities underlying the existence of predicates? [Q2] Does SEm PI effectively classify the existence of predicates? [Q3] Can the model we have trained accurately predict the existence of unknown predicates? |
| Researcher Affiliation | Academia | Man Zhu1, Pengfei Huang2, Lei Gu3, Xiaolong Xu1, Jingyu Han1 1School of Computer Science, Nanjing University of Posts and Telecommunications, China 2College of Integrated Circuits, Nanjing University of Aeronautics and Astronautics, China 3School of Computer Science and Engineering, Nanjing University of Science and Technology, China EMAIL |
| Pseudocode | Yes | Algorithm 1: Predicate Augmentation Input: G, predicate set P, ts, tc Output: predicate triples 1: cand 2: for each p in P do 3: Select the set of start relations S = {s : Ys Xr = } 4: Select the set of end relations T = {t : Xt Yr = } 5: for each possible relation sequence p1 and p2 that form rule r do 6: if supp(r)> ts and conf(r)> tc then 7: cand < p1, p2, p > 8: end if 9: end for 10: end for 11: return cand ... Algorithm 2: SEm PI Training Input: triple set T, corrupted predicate set Pcor, G Output: M, es, eo, ep and epcor 1: initialize a DNN model M 2: initialize embeddings for entities and predicates es, eo, ep and epcor 3: while not stop criterion 4: learn or update a classifier M with {ep}, and {epcor} 5: for each p P, and Pcor 6: for each < s, p, o > G 7: score+ = ||es Mp eo||2 8: end for 9: score= score ||c(M(p)) l(p)||2 10: end for 11: calculate loss function and update es, eo, ep and epcor according to the gradients 12: end while 13: return M, es, eo, ep and epcor |
| Open Source Code | No | The paper states: "We implement the whole framework based on Ampligraph (Costabello et al. 2019) using Python." This refers to a third-party library used by the authors, not their own source code for the methodology being made publicly available. There is no explicit statement of code release or a link to a repository for the work presented in this paper. |
| Open Datasets | Yes | Our experiments have been thoroughly conducted on the widely recognized FB15k and DRKG datasets, offering a robust foundation for our analysis. In table 1, we present comprehensive statistics about these datasets. |
| Dataset Splits | Yes | To assess the prediction performance with a robust methodology, we implement a 10-fold cross-validation approach on the datasets. |
| Hardware Specification | Yes | The evaluations are carried on a server using 22 v CPU AMD 64-Core processor with 90GB RAM and 1 GPU RTX 4090 with 24GB VRAM. The operation system is Ubuntu 18.04. |
| Software Dependencies | No | The paper states: "We implement the whole framework based on Ampligraph (Costabello et al. 2019) using Python." While it mentions Ampligraph and Python, it does not specify version numbers for these software components, which are crucial for reproducibility. |
| Experiment Setup | Yes | The model is a 4-layer multi-perceptron architecture with a Rectified Linear Unit (Re LU) activation function, and which is optimized with Adam optimizer. The probability parameter θ was set to 0.2 to achieve a balanced trade-off between mitigating noise issues and maintaining the learnability of the embeddings. We selected 500 predicates to corrupt in FB15k dataset, and 45 predicates to corrupt in DRKG. In the evaluations, both entity and predicate embeddings have been set to a dimension of 200. |