Source-linked AI summary
Topology-Aware Correlations Between Relations for Inductive Link Prediction in Knowledge Graphs
Jiajun Chen, Huarui He, Feng Wu, Jie Wang
TL;DR
Inductive link prediction must infer missing relations across changing entity sets, while existing approaches often overlook semantic correlations between relations. TACT models those correlations through topological patterns and a Relational Correlation Network, and experiments show significant gains over state-of-the-art inductive methods on benchmark datasets.
Problem
Existing inductive reasoning methods often focus on logical rules without modeling semantic correlations between relations, despite their prevalence in knowledge graphs.
Method
TACT categorizes relation pairs by topological structure, constructs a relational correlation graph, and uses an RCN to learn the importance of different correlation patterns for entity-independent prediction.
Results
TACT significantly outperforms several existing state-of-the-art methods on benchmark datasets for inductive link prediction.
Takeaways & Limitations
Modeling topology-aware correlations between relations effectively supports inductive link prediction.
Takeaways & Limitations
TACT takes more time than GraIL because it models correlations between relations, although its total running time remains acceptable.
Abstract
from arXiv · showhide
Inductive link prediction -- where entities during training and inference stages can be different -- has been shown to be promising for completing continuously evolving knowledge graphs. Existing models of inductive reasoning mainly focus on predicting missing links by learning logical rules. However, many existing approaches do not take into account semantic correlations between relations, which are commonly seen in real-world knowledge graphs. To address this challenge, we propose a novel inductive reasoning approach, namely TACT, which can effectively exploit Topology-Aware CorrelaTions between relations in an entity-independent manner. TACT is inspired by the observation that the semantic correlation between two relations is highly correlated to their topological structure in knowledge graphs. Specifically, we categorize all relation pairs into several topological patterns, and then propose a Relational Correlation Network (RCN) to learn the importance of the different patterns for inductive link prediction. Experiments demonstrate that TACT can effectively model semantic correlations between relations, and significantly outperforms existing state-of-the-art methods on benchmark datasets for the inductive link prediction task.
1 Introduction
Knowledge graphs are incomplete and continuously gain unseen entities, making inductive link prediction necessary for generalizing across entities. TACT addresses limitations in rule-focused approaches by modeling topology-aware semantic correlations between relations.
- Motivation: Knowledge graphs represent structured facts but remain incomplete as new entities continuously emerge in applications such as ecommerce and biomedicine.These conditions motivate predicting missing links in evolving graphs.
- Inductive link prediction: Inductive link prediction predicts missing relations when training and inference entities differ, requiring generalization to unseen entities.Many transductive methods cannot handle previously unseen entities.
- Existing approaches: Existing inductive approaches mainly learn entity-independent logical rules from relation co-occurrences or subgraph structures.Rule-learning methods explicitly mine rules, whereas GraIL learns them implicitly through subgraph reasoning.
- Core observation: Semantic correlations between relations vary, and their topological connection patterns influence those correlations.The paper contrasts strongly correlated nationality and spoken-language relations with weakly correlated nationality and film-country relations.
- TACT: TACT categorizes relation pairs by topology, builds a relational correlation graph, and uses an RCN to learn pattern-specific correlation coefficients.The resulting relation information is incorporated into entity-independent inductive link prediction.
2 Related Work
Prior inductive methods include rule-learning and embedding-based approaches, but they face limitations in modeling complex relation semantics or handling unseen graph structures. TACT contributes topology-aware relation modeling and targets inductive prediction directly.
- Rule learning: Rule-learning methods are inductive because their learned logical rules are independent of entities.However, traditional methods may scale poorly, while newer methods focus mainly on Horn rules.
- Embedding methods: Embedding-based methods can generate representations for unseen entities, but some require those entities to connect to known entities.Such requirements prevent handling entirely new graphs in the cited approaches.
- TACT contributions: TACT categorizes relation pairs into seven topological patterns and uses a relational correlation network to model topology-aware correlations.The patterns include head-to-tail, tail-to-tail, head-to-head, tail-to-head, parallel, loop, and not connected.
- Task and evaluation: TACT is designed for inductive link prediction, whereas the discussed knowledge graph embedding methods have difficulty with that setting.Its evaluation compares the approach with existing state-of-the-art inductive reasoning methods on benchmark datasets.
3 Methods
TACT performs inductive link prediction by combining topology-aware relation correlations with local graph-structure information in an entity-independent framework. It represents relation interactions through an RCG and RCN, then combines the resulting relation and structural embeddings to score triples.
- TACT scores a triple using two modules: a relational correlation module and a graph structure module, whose outputs feed a scoring network.The relational correlation module captures relation-level information, while the graph structure module represents the local graph around the target triple.
- Relational correlation module: TACT models semantic relation correlations through topological patterns and learned correlation coefficients.The method treats topological structure as informative about semantic correlations and learns the degree of different correlations.
- Relational correlation module: The RCG represents relations as nodes and their topological correlation patterns as edges, using seven patterns: H-T, T-T, H-H, T-H, PARA, LOOP, and NC.The patterns correspond to head-to-tail, tail-to-tail, head-to-head, tail-to-head, parallel, loop, and not connected configurations.
- Relational correlation network: RCN groups adjacent relations by six connected patterns, applies pattern-specific linear transformations, and aggregates their correlations into a relation neighborhood embedding.The NC pattern is excluded from these groups because it represents relations that are not connected in the knowledge graph.
- Graph structure module: The graph structure module extracts an enclosing subgraph around the target entities, labels nodes by their distances to those entities, and embeds the subgraph with R-GCN.The enclosing subgraph is formed from intersecting k-hop neighborhoods, while the target nodes receive unique distance labels.
- Scoring and training: A scoring network combines the final relational-correlation embedding with the graph-structure embedding, and negative sampling trains positive triples to score above negative triples using a hinge loss.The loss uses a margin hyperparameter and multiple negative samples for each ground-truth triple.
4 Experiments and Analysis
Experiments evaluate TACT on inductive link and relation prediction benchmarks, ablations, frequency controls, case studies, YAGO3-10, and runtime. Results support topology-aware relational correlations, while showing additional computational cost remains acceptable.
- Inductive Link Prediction: TACT-base outperforms inductive baselines on AUC-PR across datasets, while TACT provides around 4% improvement over GraIL on most datasets.These results support the effectiveness of modeling topology-aware correlations between relations.
- Inductive Link Prediction: TACT significantly outperforms rule-learning methods and GraIL for inductive relation prediction, with larger improvements on FB15k-237 and NELL-995 than WN18RR.The paper attributes this pattern to more complex relational semantics in datasets containing more relations.
- Ablation Studies: TACT outperforms TACT w/o RA and TACT w/o RC on all three benchmark datasets, supporting both relation aggregation and correlation modeling.TACT w/o RA omits neighborhood aggregation, whereas TACT w/o RC aggregates relations without modeling their correlations.
- Ablation Studies: TACT outperforms a frequency-based ranking method by a large margin, indicating its results are not explained by relation-frequency bias in the benchmark datasets.The frequency-based method returns the same frequency-ranked list for every prediction.
- Case Studies: Case studies show that TACT can learn correct correlation patterns, including assigning the largest coefficient to “has part” among neighboring relations of “member meronym.”The two relations are adjacent under the parallel topological pattern.
- Further Experiments: On YAGO3-10, which contains 37 relations and 123,182 entities, TACT outperforms GraIL on all metrics.This evaluates the method on a larger knowledge graph with relatively few relations.
- Running Time: TACT takes more time than GraIL because it models correlations between relations, but the total running time remains acceptable.The comparison is reported in the runtime analysis.
5 Conclusion
TACT exploits topology-aware correlations between relations for inductive link prediction in knowledge graphs and significantly outperforms several state-of-the-art methods on benchmark datasets.
- TACT categorizes relation pairs into topological patterns and uses a Relational Correlation Network to learn their importance for inductive link prediction.
- TACT significantly outperforms several existing state-of-the-art methods on benchmark datasets for inductive link prediction.
A. The Number of Topological Patterns
For any two irreflexive relations, topological patterns are determined by the number and arrangement of shared endpoints, yielding at most seven patterns.
- At most 7 topological patterns exist between any two irreflexive relations.
- Two relation edges can have zero, one, or two intersections, which determines their possible topological structures.
- Zero intersections produce one pattern because the two edges are not connected.
- One intersection produces four patterns: head-to-head, head-to-tail, tail-to-head, and tail-to-tail.
- Two intersections produce two patterns, corresponding to shared endpoints or reversed endpoints.
B. Additional Experiments
Additional experiments evaluate TACT on inductive relation prediction, scoring-network embeddings, and relation-correlation examples. TACT outperforms competing methods, while different embedding components contribute to final performance.
- Inductive Relation Prediction: TACT significantly outperforms rule-learning methods and GraIL on MRR and H@1 across inductive benchmarks from WN18RR, FB15k-237, and NELL-995.The improvements are larger on FB15k-237 and NELL-995, which contain more relations and more complex semantic correlations.
- Inductive Relation Prediction: TACT models complex relational patterns by exploiting correlations between relations, whereas GraIL struggles when the number of relations is large.
- Scoring Network: The scoring network uses final relation, graph, and node embeddings, and experiments evaluate different combinations of these embedding components.
- Scoring Network: Each embedding component contributes to final results, while using only the final relation embedding provides fairly good performance for inductive relation prediction.
- Case Studies: Case studies present the top three relevant relations for selected relations from WN18RR and NELL-995 using correlation patterns and coefficients.