PaperlayerView exact revision on arXiv

Source-linked AI summary

PEARL: Path-Entity Aligned Relational Learning with Contextual Subgraphs for Inductive Knowledge Graph Completion

Yunchi Yang, Longlong Li, Cunquan Qu

arXiv:2609.02216v1cs.AI

TL;DR

IKGC must predict missing links for entities unseen during training, yet path relevance depends on surrounding structural context. PEARL retrieves semantically relevant paths from query-specific contextual subgraphs, aligns them with entities and global structure, and regularizes representations with dual-view contrastive learning. It achieves the best average Hits@10 among compared methods on WN18RR, FB15k-237, and NELL-995.

  • Problem

    IKGC requires transferable relational and structural patterns for unseen entities, while existing approaches may encode paths independently of query-subgraph context despite context-dependent path relevance.

  • Method

    PEARL builds union-based contextual subgraphs, retrieves semantically relevant paths with an LLM-guided module, aligns paths with contextual entities and global structure, and applies contrastive regularization.

  • Results

    PEARL achieves the best average Hits@10 among compared IKGC methods on WN18RR, FB15k-237, and NELL-995.

  • Takeaways & Limitations

    Jointly exploiting structural context and relational paths provides an effective paradigm for inductive knowledge graph completion.

Abstract

from arXiv · show

Inductive knowledge graph completion (IKGC) aims to predict missing links involving entities unseen during training, requiring models to learn transferable relational and structural patterns. Existing subgraph- and path-based approaches often encode relational paths independently of their surrounding query subgraphs, although the predictive relevance of a path may vary across structural contexts. We propose PEARL, a Path-Entity Aligned Relational Learning framework that models paths as context-conditioned reasoning signals. PEARL constructs a query-specific contextual subgraph from the union of the query entities' neighborhoods and uses a large language model (LLM)-guided retriever to distill semantically relevant paths. It then builds a bipartite interaction graph over paths, contextual entities, and a global subgraph representation, allowing path embeddings to adapt to local and global structural evidence. To suppress noise introduced by the enlarged context, PEARL employs a dual-view contrastive objective that promotes representation consistency under stochastic contextual perturbations. Experiments on WN18RR, FB15k-237, and NELL-995 show that PEARL obtains the best average Hits@10 among the compared IKGC methods on all three benchmarks. Ablation studies, efficiency analyses, and case studies further validate the contributions of contextual subgraph modeling, semantic path retrieval, path-entity interaction, and contrastive regularization.

1 Introduction

Inductive knowledge graph completion must generalize relational and structural patterns to unseen entities, while PEARL addresses context-dependent path relevance and noisy evidence through aligned contextual reasoning.

  • Inductive knowledge graph completion requires transferable relational and structural patterns because training and test entities are disjoint.
  • Existing methods use rule-based, subgraph, or path evidence, but relational paths may vary in predictive value across structural contexts.
  • PEARL constructs union-based contextual subgraphs, applies dual-view contrastive learning, retrieves informative paths, and aligns them with contextual entities and global structure.
  • PEARL jointly models contextual subgraphs and relational paths to improve reasoning over unseen entities.
  • LLM-guided retrieval and path–entity alignment identify informative relational evidence and learn robust context-aware representations.
  • PEARL achieves the best average Hits@10 across WN18RR, FB15k-237, and NELL-995 among the compared methods.

2 Related Work

Related work for inductive knowledge graph completion spans rule-based, subgraph-based, and path-based reasoning, with PEARL integrating these structural and relational perspectives contextually.

  • Inductive knowledge graph completion predicts missing triples involving unseen entities and requires transferable relational patterns rather than entity-specific embeddings.
  • Rule-based methods: Rule-based methods learn logical rules that support inductive inference, including differentiable reasoning and variable-length rule modeling.
  • Subgraph-based methods: Subgraph-based methods encode local structures around query entities, with later approaches improving relation awareness, topology, neighborhood quality, and multi-view modeling.
  • Path-based methods: Path-based methods capture explicit multi-hop relation sequences connecting query entities and provide complementary evidence for inductive reasoning.
  • Subgraph & path-based methods: PEARL combines contextual subgraphs, LLM-guided path retrieval, path–entity interactions, contrastive learning, and path-attention fusion in one framework.

3 Preliminary

The preliminary formulation defines knowledge graphs, query-local neighborhoods, contextual subgraphs, and inductive link prediction through transferable scoring patterns.

  • Knowledge graph: A knowledge graph is G = (E, R, T), where entities and relation types form directed relational triples.
  • Contextual subgraph: For a query triple, subgraph-based IKGC uses undirected k-hop neighborhoods around query entities to define local structural context.
  • Contextual subgraph: Prior work uses an enclosing subgraph, while PEARL obtains a contextual subgraph from the union of the query endpoints’ neighborhoods.
  • Contextual subgraph: The enclosing subgraph is contained within the contextual subgraph, with the former induced by its enclosing-subgraph vertices.
  • Inductive link prediction: Inductive link prediction learns a scoring function that ranks positive triples above corrupted triples without relying on entity-specific identifiers.

4 Methodology

PEARL jointly models contextual subgraphs and relational paths for inductive link prediction, using semantic path retrieval, path–entity alignment, and contrastive learning to preserve useful context while suppressing noise.

  • Contextual Subgraph Modeling: PEARL constructs a query-specific contextual subgraph and encodes contextual entities plus a global subgraph representation with CSNN.The contextual subgraph provides richer structural evidence for reasoning over unseen entities.
  • Contrastive Regularization: A dual-view subgraph contrastive objective promotes consistency across augmented subgraphs and suppresses unstable contextual noise.The objective encourages representations to focus on stable, query-relevant structural semantics.
  • Contextual Subgraph Modeling: CSNN uses structural and relational features instead of fixed entity identifiers, supporting inductive representations for unseen entities.Node features include double-radius labels and query-aware aggregation of incident relation embeddings.
  • LLM-based Path Retrieval and Alignment: PEARL retrieves candidate paths with an LLM, ranks their semantic compatibility with the query triple, and retains the most informative paths.Candidate paths are first extracted by breadth-first search, while path counts increase from 2.7 in V1 to 15.1 in V4 on FB15k-237.
  • Path–Entity Interaction: Retained paths are aligned with contextual entities and a global node in a bipartite graph containing path–entity and path–global connections.This graph explicitly represents correspondences between relational paths, local entities, and global contextual semantics.
  • Path–Entity Interaction: BGNN with type-aware R-GAT message passing refines each path embedding using surrounding entity information and global subgraph semantics.Consequently, identical relational paths can be interpreted differently in different structural environments.

5 Experiments

PEARL is evaluated on three inductive benchmarks using standard filtered ranking, and it achieves the strongest average performance while maintaining favorable efficiency. Ablations and case studies support the roles of contextual subgraphs, contrastive learning, semantic path retrieval, and path–entity interaction.

  • Results and Analysis: PEARL achieves relative average Hits@10 improvements of 8.67%, 6.01%, and 3.20% over the strongest competitors on WN18RR, FB15k-237, and NELL-995, respectively.
  • Results and Analysis: PEARL jointly models contextual subgraphs and relational paths, using LLM-guided retrieval and BGNN interactions to combine structural and logical evidence.
  • Efficiency Analysis: PEARL improves rapidly during early training, converges faster than most baselines, and attains the highest validation AUC-PR on FB15k-237 V1 and V3.It also requires the lowest training GPU memory on V1 and remains comparable to the most memory-efficient baseline on V3.
  • Efficiency Analysis: LLM-guided retrieval adds preprocessing overhead but removes redundant paths, enabling competitive downstream inference efficiency and the best Hits@10 on both inductive settings.
  • Ablation on Subgraph Modeling: Contextual subgraphs outperform conventional enclosing subgraphs, while contrastive learning mitigates noise introduced by incorporating entities from the union of query neighborhoods.
  • Ablation on Path-aware Reasoning: Removing paths, replacing semantic retrieval with random selection, or removing BGNN degrades performance; BGNN removal causes the largest drop.Random selection also produces larger standard deviations, while paths provide complementary evidence beyond contextual subgraphs.
  • Case Study: For a representative query, the path “Film release region →location contains →film release region” receives the highest importance score because it directly captures the required geographical relation.Other candidate paths receive substantially lower scores because they provide indirect or less relevant evidence.
  • Case Study: Path importance depends jointly on relation sequence and surrounding contextual subgraph, allowing PEARL to interpret paths according to local structural context.

6 Conclusion

The paper presents PEARL as a context-aware framework for inductive reasoning over unseen entities. Experiments show strong overall performance, supporting the joint use of structural context and relational paths.

  • PEARL jointly models contextual subgraphs and relational paths for inductive reasoning over unseen entities.
  • Experiments on WN18RR, FB15k-237, and NELL-995 demonstrate strong overall performance against competitive baselines.
  • The results support jointly exploiting structural context and relational paths for inductive knowledge graph completion.

Supplemental Materials

PEARL uses an LLM to rank candidate relational paths by semantic relevance to each query triple. The resulting ranking prioritizes paths for subsequent reasoning.

  • PEARL uses Qwen3-4B-Instruct to assign importance scores to candidate paths according to their semantic relevance to the query triple.
  • Candidate paths are ranked by their importance scores, and the ranking prioritizes informative reasoning evidence for subsequent processing.

B Dataset and Evaluation Protocol

Evaluation uses three IKGC benchmarks divided into inductive subsets with disjoint training and test entities. Performance is measured with filtered ranking metrics and validation AUC-PR for checkpoint selection.

  • PEARL is evaluated on WN18RR, FB15k-237, and NELL-995, three widely used IKGC benchmarks.
  • Each benchmark is divided into four increasing-scale inductive subsets, with disjoint training and test entity sets sharing a global relation vocabulary.
  • Validation AUC-PR is used to select the checkpoint for final testing.

C Details of Baseline Methods

The baselines span differentiable rule learning, enclosing-subgraph reasoning, topology-aware message passing, and relational-path contrastive methods for inductive link prediction.

  • NeuralLP learns logical rules for multi-hop reasoning through an end-to-end differentiable TensorLog-based framework.
  • DRUM extends differentiable rule learning to variable-length logical rules by sharing parameters across path lengths.
  • GraIL performs inductive knowledge graph completion by reasoning over enclosing subgraphs.
  • Several subgraph methods enhance local reasoning through directional relations, topology-aware relation correlations, relational message passing, or locality-aware neighborhood extraction.These include CoMPILE, TACT, RMPI, and LCILP.
  • Other approaches balance structural views, enable neighbor message intercommunication, aggregate differentiable paths, capture unseen-entity semantics, or strengthen paths contrastively.These include S2DN, MINES, NBFNet, SNRI, and RPC-IR; CATS1 additionally uses an LLM with latent type constraints, paths, and neighboring facts.

D Further Analysis of Contextual Subgraph Modeling

The contextual-subgraph analysis finds that contrastive learning improves alignment between richer contextual representations and core enclosing-subgraph semantics while preserving discriminative information.

  • Without SCL, contextual subgraph embeddings deviate more from the core structural semantics captured by enclosing subgraphs.
  • PEARL with SCL achieves better alignment while retaining richer contextual information, suggesting reduced influence from noisy contextual perturbations.
  • Positive Consistency and Separation evaluate representation quality for contextual and enclosing subgraph embeddings within a mini-batch.The setup uses B query triples and compares the two embedding views.
  • PEARL achieves higher Cpos and S than PEARL w/o SCL, indicating improved semantic consistency and preserved discriminative contextual representations.
  • Figure 6 compares global subgraph embeddings for PEARL w/o CS, PEARL w/o SCL, and PEARL.

E Hyperparameter Sensitivity

PEARL is generally stable across batch sizes, but optimal learning rates, receptive-field sizes, path counts, and contrastive weights vary by dataset.

  • PEARL remains relatively stable across batch sizes from 8 to 64 on all three datasets.The best batch size is 8 for WN18RR and NELL-995, versus 32 for FB15k-237.
  • Excessively large learning rates slightly degrade performance; the best values are 0.001 for WN18RR and NELL-995 and 0.0005 for FB15k-237.
  • The best subgraph hop size is k = 4 on WN18RR, k = 3 on FB15k-237, and k = 2 on NELL-995.The preferred structural receptive field therefore depends on dataset characteristics and density.
  • The best retained path count is M = 3 on WN18RR and FB15k-237, while NELL-995 performs best with M = 5.
  • The study varies λ1 and λ2 to measure how task loss and subgraph contrastive loss contributions affect Hits@10 on the three V1 datasets.
  • WN18RR favors a relatively small contrastive weight, whereas FB15k-237 and NELL-995 benefit from larger λ2 values.

F Impact of Relation Frequency

Relation-frequency analysis shows that PEARL remains effective across low-, medium-, and high-frequency query-relation groups, with especially strong results in several higher-frequency regimes.

  • Test triples are partitioned into bottom 25%, middle 50%, and top 25% groups according to query-relation frequency in the training set.
  • PEARL achieves the best or competitive performance across most relation-frequency groups on WN18RR, FB15k-237, and NELL-995.
  • PEARL’s gains are particularly evident in WN18RR’s high-frequency group and FB15k-237’s medium- and high-frequency groups.
  • On NELL-995, PEARL performs best in the low- and high-frequency groups while remaining competitive in the middle-frequency group.
  • The analysis concludes that PEARL is robust across relation-frequency regimes and can exploit relational and structural evidence under diverse frequency distributions.
  • Table 8 reports Hits@10 percentages for relation-frequency groups, highlighting the best results in bold.
Loading 2609.02216v1…