Collective Entity Alignment via Adaptive Features

I. INTRODUCTION

• Entity alignment identifies equivalent entities across knowledge graphs and is an indispensable step in incorporating target-graph knowledge into a source graph.The motivation is that knowledge graphs do not reach full coverage and often contain complementary content.Source: §I · p. 1
• Embedding-based methods model structural features and rank target entities separately for each source entity.The top-ranked target is selected as the match.Source: §I · p. 1
• Independent entity decisions overlook interdependence: a target is less likely to match a source if it is aligned to another source with higher confidence.The authors identify collective signals in decision-making as distinct from using structural information during feature generation.Source: §I · p. 1
• CEA combines structural, semantic, and string-level features, then formulates alignment as stable matching solved by deferred acceptance.The framework was evaluated on cross-lingual and mono-lingual tasks against 11 state-of-the-art methods.Source: §I · p. 1

II. PROBLEM DEFINITION AND RELATED WORK

• EA seeks equivalent target entities for source entities in a test set, given two knowledge graphs and seed entity pairs.The seed pairs encode known equivalences between entities in the source and target graphs.Source: §II · p. 1
• Structural-information approaches encode graph structure in low-dimensional spaces and align those spaces using seed entity pairs.Some methods instead fuse training corpora and project entities directly into a shared embedding space.Source: §II · p. 1
• Prior multi-source approaches use attributes, descriptions, and entity names, whereas CEA aggregates features at the outcome level rather than the representation level.The cited sources include attribute types [11],, attribute values [13], and entity names [15], [17], [18].Source: §II · p. 2

III. FEATURE GENERATION

• CEA adapts three generally available feature types to address entity alignment.They are structural, semantic, and string-level signals.Source: §III · p. 2
• GCN encodes entity neighborhood information as real-valued structural vectors, whose pairwise similarity is calculated with cosine similarity.The resulting structural similarity matrix has source entities as rows and target entities as columns.Source: §III · p. 2
• Averaged word embeddings represent entity-name semantics without requiring a special training corpus, and cosine similarity forms the semantic similarity matrix.The name embeddings are represented in matrix form as N; the semantic similarity matrix is denoted Mn.Source: §III · p. 2
• Levenshtein distance supplies string similarity, which is useful for mono-lingual and closely related cross-lingual KGs without external resources or out-of-vocabulary effects.The corresponding string similarity matrix is denoted Ml.Source: §III · p. 2

IV. COLLECTIVE ENTITY ALIGNMENT

• Equal-weighted feature similarity matrices are fused into M before the matching procedure.The independent alternative ranks targets by each source entity’s row in M and chooses the top-ranked target.Source: §IV · p. 2
• EA is formulated as stable matching because independent ranked matches fail to model interdependence between alignment decisions.Preference lists rank entities on the opposing side by descending similarity scores.Source: §IV · p. 2
• Fig. 1 presents entity alignment as stable matching using deferred acceptance.The accompanying example follows proposals among three source and three target entities using the fused similarity matrix.Source: Figure 1 · p. 2, §IV · p. 2
• In deferred acceptance, unmatched sources propose to their highest-ranked untried targets, which retain a preferred proposal and may reject a provisional match.Rounds continue until every source entity is settled.Source: §IV · p. 2
• The procedure guarantees every entity is matched and the resulting marriages are stable.Stability means no source-target pair would both prefer each other over their assigned partners.Source: §IV · p. 3
• Stable matching is presented as more desirable than maximum-weight bipartite matching because it avoids the Hungarian algorithm’s stated O(n^4) computational cost and yields matchings without incentives to deviate.The paper describes maximum-weight bipartite matching as a possible alternative.Source: §IV · p. 3

Evaluation results

• CEA consistently outperforms all baselines in the reported evaluation results.The comparisons are reported for DBP15K and SRPRS.Source: Table I · p. 3, Table II · p. 3, §V · p. 3

Structural methods

• On SRPRS, RSNs exceed BootEA, reversing their ordering on DBP15K; RSNs model long-term relational dependencies, while BootEA uses one-to-one constrained bootstrapping.The passage attributes RSNs’ SRPRS result to real-life degree distributions and their reduced sensitivity to entity degrees.Source: §V · pp. 3–4

Methods using multiple information sources

• CEA outperforms name-based methods by a large margin, aggregating features at outcome level rather than representation level.Attribute-based methods show dataset-dependent results, while entity-name information improves several competing methods.Source: §V · p. 4

Evaluation results

• On mono-lingual datasets, CEA reaches almost 1.000 accuracy; the authors attribute this to nearly identical entity names and effective string-level features.The authors suggest more challenging mono-lingual EA datasets as future work.Source: §V · p. 4

Ablation study

• Without collective EA, performance drops on all tested datasets, supporting the value of accounting for interdependence between alignment decisions.The ablation results cover SRPRS and DBP15KZH-EN.Source: §V · p. 4, Table III · p. 4
• Structural information consistently helps cross-lingual alignment; semantic features matter more for distant language pairs, while string features matter more for closely related pairs.On mono-lingual datasets, removing structural or semantic information does not hurt accuracy, but removing string-level features causes a considerable drop.Source: §V · p. 4

• CEA combines adaptive structural, semantic, and string features with stable matching for collective entity alignment.Source: §VI · p. 4
• The framework addresses existing EA methods' overlooked opportunity to improve performance through a collective alignment solution.Source: §VI · p. 4
• Experiments verify that the resulting method is superior to state-of-the-art options.Source: §VI · p. 4