Research map / Graph learning / Knowledge and temporal graphs
Representation Learning Based Predicate Invention on Knowledge Graphs
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the ReLPI problem and SEmPI, a framework that learns diagonal-matrix predicate embeddings with a scoring function and an MLP classifier to decide whether new predicates should be invented in knowledge graphs.
Problem
Deciding whether a new predicate should be invented in a knowledge graph, using the statistical patterns of existing predicate occurrences rather than logic-based ILP methods that scale poorly.
Approach
Encodes predicates as diagonal matrices so that composition (matrix multiplication) for transitive rules, averaging for conjunctive rules, and inversion for reverse rules approximate candidate predicate embeddings; a loss combining a classifier term and a subject-matrix-object distance term is optimized jointly with a DNN invent/non-invent classifier.
Key results
- On FB15k, the trained SEmPI classifier reaches 0.820 precision, 0.724 recall, 0.674 accuracy and 0.779 F1 on 610 candidate predicates, judged by artificial (manual) evaluation.
- SEmPI is reported to achieve higher recall than TransE-DNN, DistMult-DNN and ComplEx-DNN in 10-fold cross-validation, at a slight cost in precision.
- Per-epoch training time for SEmPI is 1528 ms on FB15k and 9890 ms on DRKG, comparable to the embedding baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SEmPI instead of TransE-DNN, DistMult-DNN, or ComplEx-DNN | Predicate existence classification · FB15k and DRKG | recall | Helped |
| Use SEmPI instead of TransE-DNN, DistMult-DNN, or ComplEx-DNN | Predicate existence classification · FB15k and DRKG | precision | Hurt |
Surprising findings
- Recall gains come at a slight precision cost, since some newly discovered predicates are not aligned with the dataset.
Limitations
- Evaluation of invented predicates is partly manual (artificial evaluation) and covers only 610 candidates of length 2 on FB15k.
- Negative examples are synthetic corruptions of triples, which may not reflect real non-existing predicates.
- Only a limited set of baselines is compared, and the ILP systems were judged effective only on logic-rich small datasets.
Experimental setup
Datasets: FB15k DRKG
Benchmarks: FB15k DRKG
Metrics: precision recall F1 accuracy training time per epoch
Baselines: TransE-DNN DistMult-DNN ComplEx-DNN Metagol HEXMIL Popper
Builds on: TransE DistMult ComplEx Popper
Closest papers by meaning
- Toward a Unified Statistical Theory of Unsupervised Pretraining and Supervised Neural Knowledge Graph Learning (NeurIPS 2026 · similarity 0.85)
- Interpreting Knowledge Graph Relation Representation from Word Embeddings (ICLR 2021 · similarity 0.85)
- Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.84)
- Reasoning of Large Language Models over Knowledge Graphs with Super-Relations (ICLR 2025 · similarity 0.83)
- Talk like a Graph: Encoding Graphs for Large Language Models (ICLR 2024 · similarity 0.83)
- Towards Foundation Models for Knowledge Graph Reasoning (ICLR 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.