Source-linked AI summary
Multi-Channel Graph Neural Network for Entity Alignment
Yixin Cao, Zhiyuan Liu, Chengjiang Li, Zhiyuan Liu, Juanzi Li, Tat-Seng Chua
TL;DR
Entity alignment must cope with heterogeneous KG structures and limited seed alignments. MuGNN jointly completes and prunes KGs through rule transfer and multi-channel attention, then learns alignment-oriented embeddings; experiments report a 5% average Hits@1 improvement across five datasets.
Problem
Entity alignment remains difficult because KGs differ structurally and available seed alignments are limited.
Method
MuGNN jointly infers and transfers rules for KG completion and uses self-attention and cross-KG attention in pooled GNN channels for alignment-oriented embeddings.
Results
5% Hits@1 improvement on average is reported across five publicly available entity-alignment datasets.
Takeaways & Limitations
MuGNN alleviates negative impacts from structural heterogeneity and limited seed alignments through completion, pruning, and robust graph encoding.
Takeaways & Limitations
The inferred rules do not hold in all cases, and confidence values for individual groundings are left for future work.
Abstract
from arXiv · showhide
Entity alignment typically suffers from the issues of structural heterogeneity and limited seed alignments. In this paper, we propose a novel Multi-channel Graph Neural Network model (MuGNN) to learn alignment-oriented knowledge graph (KG) embeddings by robustly encoding two KGs via multiple channels. Each channel encodes KGs via different relation weighting schemes with respect to self-attention towards KG completion and cross-KG attention for pruning exclusive entities respectively, which are further combined via pooling techniques. Moreover, we also infer and transfer rule knowledge for completing two KGs consistently. MuGNN is expected to reconcile the structural differences of two KGs, and thus make better use of seed alignments. Extensive experiments on five publicly available datasets demonstrate our superior performance (5% Hits@1 up on average).
1 Introduction
Entity alignment is challenged by heterogeneous KG structures and insufficient seed alignments. MuGNN jointly uses completion, pruning, and multi-channel encoding to reconcile these differences and improve alignment.
- KGs contain entities and relations in directed-graph form and often provide complementary structural knowledge for applications.
- Structural heterogeneity can make incorrect entity matches appear more similar and propagate alignment errors across the graph.The Jilin City example contrasts informative and missing relations across English and Chinese KGs.
- Limited seed alignments constrain embedding quality, while GCN-based approaches may still fail to account for structural heterogeneity.
- MuGNN reconciles missing relations and exclusive entities through channels targeting KG completion and pruning, combined with pooling.
- MuGNN jointly performs KG inference and alignment, using rule inference and transfer plus cross-KG attention to reconcile heterogeneity.
- 5% Hits@1 improvement on average is reported across five publicly available entity-alignment datasets.
2 Preliminaries and Framework
The framework represents heterogeneous KGs as directed relational graphs and learns alignment-oriented embeddings through shared, multi-channel graph encoding. Rule-based completion and relation-weighted GNN channels feed an alignment model.
- 2.1 Preliminaries: A KG is modeled as G = (E, R, T), with entities, relation types, and relational triplets forming a directed graph.
- 2.1 Preliminaries: Entity alignment seeks equivalent entities across two KGs, using manually or lexicon-derived seed alignments as prior correspondences.
- 2.2 Framework: MuGNN first completes missing KG relations by inducing rules and transferring them between graphs based on aligned relations.
- 2.2 Framework: The multi-channel GNN applies self-attention and cross-KG attention to weight relations, shares channel parameters, and pools channel outputs.
- 2.2 Framework: The Align model maps both KGs into a unified vector space by bringing seed-aligned entities and relations together.
3 KG Completion
KG completion mines rules within each graph, transfers them across aligned relations, and grounds them to add missing conclusions. The method assumes knowledge can generalize across KGs, but inferred rules are not universally valid.
- KG completion mines rules separately from each KG, transfers them across graphs, and grounds them for consistent completion.
- Rule transfer relies on the Knowledge Invariant Assumption that knowledge has universality across languages or domains.
- Transferred rules replace relations with aligned counterparts and are added to the target rule set when absent.
- Rule grounding identifies premise-satisfying cases lacking conclusion triplets, then adds those conclusions to each KG.
- A transferred province-and-dialect rule completes a missing dialect relation for Jilin City in KG1.
- Inferred rules do not hold in all cases; assigning confidence values to individual groundings is left for future work.
4 Multi-Channel Graph Neural Network
MuGNN encodes each knowledge graph through relation-weighted, multi-channel GNNs that separately address missing relations and exclusive entities. The resulting embeddings are aligned using seed information while rule-based completion is incorporated into the training objective.
- Relation weighting: MuGNN generates two weighted adjacency matrices per KG: KG self-attention for completion-oriented propagation and cross-KG attention for pruning-oriented propagation.The two channels reconcile missing relations and exclusive entities separately.
- KG self-attention: KG self-attention assigns higher weights to informative neighbors using the current entity and the KG’s own structure.The normalized connectivity aij is computed from attention coefficients, whose parameters include W and p.
- Cross-KG attention: Cross-KG attention models common subgraphs by assigning lower relation weights when corresponding relation types lack counterparts in the other KG.Its similarity measure is the inner product between relation representations, and the resulting weight is zero when no matching relation type exists.
- Multi-channel GNN encoder: Each GNN layer propagates neighbor-informed representations through channel-specific connectivity matrices, then combines channel outputs using pooling.The encoder stacks multiple layers; experiments select average pooling because it performs best.
- Align model: The align model embeds both KGs in a unified vector space, while triplet loss preserves newly added grounded rules as valid.Alignment uses L2 distance with margin hyper-parameters, and negative samples are periodically recalculated.
5 Experiment
Experiments on five public datasets evaluate MuGNN against translation- and graph-based baselines, with overall results, ablations, seed-sensitivity tests, and qualitative rule-transfer analysis. MuGNN outperforms baselines broadly, while its gains depend on structural reconciliation, cross-KG attention, rule availability, and seed proportion.
- Overall Performance: MuGNN significantly outperforms all baselines across evaluation metrics on DBP15K and DWY100K.The comparison uses 30% of entity seed alignments for training, with Hits@N and MRR as evaluation metrics.
- Overall Performance: 3%, 6%, and 4% are MuGNN’s average gains in Hits@1, Hits@10, and MRR on three small-scale cross-lingual datasets.On large-scale datasets, the corresponding gains are 8%, 8%, and 8%.
- Impact of Two Channels and Rule Transfer: Removing either attention channel reduces performance, with cross-KG attention causing the larger decrease.The ablation compares MuGNN without KG self-attention, cross-KG attention, or rule transfer.
- Impact of Two Channels and Rule Transfer: Rule transfer contributes substantially in most cases, but its effect is negligible on DBPFR-EN because ground rule triples are less than 1% of oracle triples.DBPJA-EN and DBPZH-EN each provide more than 10k ground-rule triples and obtain decent improvements from rule transfer.
- Impact of Seed Alignments: At 10% seed alignments, GNN-based methods outperform translation-based methods, but their performance worsens as seed alignments increase.MuGNN explicitly completes KGs through AMIE+ rule mining and transfer while using two GNN channels to exploit seed information.
- Qualitative Analysis: Transferred rules smooth asymmetric information by adding missing knowledge across KGs, producing more similar structures for corresponding entities.Examples transfer common rules between English and Chinese DBpedia graphs in both directions.
6 Related Work
Prior entity-alignment methods use external features, iterative seed expansion, or graph encoders, but limited seeds and structural differences remain central challenges. MuGNN instead targets structural heterogeneity through completion and pruning without external information or iterative alignment.
- Earlier approaches use machine translation, external lexicons, attributes, or descriptions to address heterogeneity between knowledge graphs.
- IP-TransE and BootEA iteratively enlarge the seed set by selecting predicted alignments with high confidence.
- General enhancement strategies motivate MuGNN’s focus on alignment without external information and in a non-iterative way.
- MuGNN reconciles heterogeneous KGs through completion and pruning while learning alignment-oriented embeddings with graph neural networks.
7 Conclusions
MuGNN learns alignment-oriented KG embeddings by using multiple relation-weighting channels to complete KGs and prune exclusive entities. Experiments on five public datasets and further analysis demonstrate its effectiveness, while future work considers entity text and cross-KG proximity.
- MuGNN learns alignment-oriented KG embeddings for entity alignment through robust multi-channel graph encoding.
- Its KG self-attention and cross-KG attention channels complete KGs and prune exclusive entities using different relation-weighting schemes.
- Experiments on five publicly available datasets and further analysis demonstrate the effectiveness of MuGNN.
- Future work will introduce entity text while considering word ambiguity and use cross-KG entity proximity for alignment.