PaperlayerView exact revision on arXiv

Source-linked AI summary

GRAND-HC: Graph-Refined Author Name Disambiguation

Yuanhao Sun, Zhouyang Jin, Yi Xu, Luoyi Fu, Jiaxin Ding, Xiaoying Gan, Xinbing Wang, Chenghu Zhou

arXiv:2609.01636v1cs.IR

TL;DR

SND must separate papers sharing an author name despite long-tailed author distributions and unreliable cluster-number estimation. GRAND-HC addresses these issues with graph-based representation, reweighted contrastive learning, distance refinement, and paper compression, achieving state-of-the-art macro-F1 performance and billion-scale deployment.

  • Problem

    SND methods face prolific-author bias that over-merges tail authors and cluster-number estimators that become unreliable for long paper sequences.

  • Method

    GRAND-HC constructs a heterogeneous paper graph and combines HCL, GRDM, and PCM before hierarchical agglomerative clustering.

  • Results

    GRAND-HC achieves state-of-the-art macro-F1 performance among recent SND methods and has been deployed in a billion-scale academic system.

  • Takeaways & Limitations

    The framework targets scalable real-world author disambiguation by addressing representation imbalance and long-sequence cluster estimation within one pipeline.

  • Takeaways & Limitations

    Distance-based estimators rely on hand-crafted thresholds, while RNN-based estimators struggle with overlong sequences and cannot capture global graph structure as paper counts grow.

Abstract

from arXiv · show

From-Scratch Name Disambiguation (SND) groups papers sharing an ambiguous name into clusters of distinct real-world authors. Existing methods suffer from two critical limitations: (1) inherent long-tailed author distribution biases representation learning, causing over-merging of tail authors; (2) existing cluster number estimation methods are unreliable for long paper sequences, hindering large-scale deployment. We propose \textbf{GRAND-HC}, a complete end-to-end SND framework. We construct a heterogeneous paper graph via co-author, co-organization, and co-venue relations, using a graph attention network as the embedding backbone. \textbf{Harmony Contrastive Learning (HCL)} dynamically reweights training loss to suppress overfitting to prolific authors, learning discriminative embeddings. A \textbf{Graph-Refined Distance Matrix (GRDM)} leverages graph topology to optimize pairwise distances, further preventing tail author over-merging. Meanwhile, a lightweight \textbf{Paper Compression Module (PCM)} achieves accurate cluster number estimation across varying scales. Finally, Hierarchical Agglomerative Clustering outputs the final clusters. Extensive experiments demonstrate state-of-the-art macro F1 performance. GRAND-HC has been deployed in a billion-scale academic database. Source code: https://github.com/baokou-fw2/GRAND-HC.

Introduction

From-Scratch Name Disambiguation groups papers sharing an ambiguous name by their distinct real-world authors, supporting academic information systems. GRAND-HC targets long-tailed representation bias and unreliable cluster-number estimation with an end-to-end framework.

  • Problem: SND partitions papers associated with one ambiguous name into disjoint clusters, each corresponding to a distinct author.The task supports academic knowledge graph construction, scholar evaluation, and academic search.
  • Long-tailed distribution: Long-tailed paper distributions bias representation learning toward prolific authors, causing tail-author papers to overlap without clear clustering boundaries.The resulting low-discriminative embedding space contributes to tail-author over-merging.
  • Long-tailed distribution: Nearly 40% of ambiguous names contain a prolific author responsible for more than 50% of papers, while 29.1% have author-count standard deviations above 50.These statistics come from the WhoisWho-SND and AMiner-v2 benchmarks.
  • Cluster number estimation: Distance-based and RNN-based estimators underpredict cluster numbers, averaging 27 and 46 clusters versus the ground-truth average of 62.On AMiner-v2, the reported values are for distance-based and RNN-based methods, respectively.
  • Cluster number estimation: As paper counts increase from 100 to 900, RMLSE rises from 0.96 to 3.72 for distance-based methods and from 0.07 to 1.24 for RNN-based methods.The result demonstrates worsening estimation error at larger input scales.
  • GRAND-HC: GRAND-HC combines Harmony Contrastive Learning, a Graph-Refined Distance Matrix, and a Paper Compression Module to address representation bias and cluster-number estimation.The framework is designed as a unified end-to-end SND pipeline.

Related Work

Related work frames SND as clustering based on paper representation learning and cluster-number estimation. The paper defines the task as partitioning candidate papers for an ambiguous name into author-specific subsets.

  • SND methods: SND methods typically learn paper representations, compute similarities, and cluster papers into groups corresponding to real authors.Representation learning approaches include feature-based and graph-based methods.
  • Representation learning: Prior work combines paper content with relational information such as co-authors, institutions, citations, heterogeneous networks, and adversarial or semi-supervised learning.Examples include semantic fingerprints, linkage models, adversarial representation learning, and heterogeneous information networks.
  • Cluster-number estimation: Because the true author count is often unknown, cluster-number estimation is crucial, while pairwise relationship methods may not directly produce clusters and can be computationally inefficient.Existing estimation methods are discussed as a separate part of the SND literature.
  • Problem formulation: A paper is represented by multiple attributes, including title, abstract, keywords, authors, organizations, venue, and publication date.The paper denotes these attributes as x_f, with F attributes in total.
  • Problem formulation: Given candidate papers containing an ambiguous author name, SND partitions them into disjoint subsets so each subset contains papers from one author.The number of subsets K equals the number of distinct authors sharing the name.

Methodology

GRAND-HC builds a heterogeneous paper graph, learns embeddings with harmony contrastive learning, refines distances using graph structure, estimates cluster size, and applies HAC.

  • Graph construction: The framework constructs a heterogeneous paper relationship network for candidate papers, with paper nodes carrying attributes for semantic extraction.Connected components represent potential article sets belonging to individual authors.
  • Graph construction: Graph edges encode co-author, co-organization, and co-venue relationships between papers.Co-author links exclude the target author; co-organization links use shared organizations; co-venue links use shared journals or conferences.
  • Attribute similarity: The method uses text-matching techniques to measure similarity across paper attributes.The detailed matching procedures are introduced for subsequent edge construction.
  • Relation matching: Co-author matching normalizes name formatting, segmented matching handles name-order variation, and edit distance corrects abbreviations and minor spelling errors.These operations address inconsistencies in co-author identification.
  • Relation matching: Co-organization matching uses the Jaccard index to handle granularity differences across universities, departments, and geographic levels.The procedure first identifies the target author’s affiliation.

Paper Embedding Generation

GRAND-HC generates paper embeddings by combining BERT semantics with heterogeneous graph information, then trains the representation network with Harmony Contrastive Learning to reduce prolific-author bias and improve author separation.

  • Embedding Inputs: BERT encodes paper titles, abstracts, venues, authors, organizations, and keywords into 768-dimensional initial semantic embeddings.
  • Graph-Semantic Fusion: Connected-component labels encode graph structure, while four independent MLPs align semantic and structural inputs into a common dimension.The four inputs include BERT embeddings, subgraph labels, degree, and year.
  • Graph-Semantic Fusion: A transformer encoder and a GAT-based structure encoder combine the aligned features to produce final paper embeddings and a cosine similarity matrix.
  • Motivation: Long-tailed paper distributions cause standard contrastive learning to overfit prolific authors, leaving tail-author papers overlapped and prone to over-merging.
  • Harmony Contrastive Learning: HCL dynamically identifies hard pairs through HAC predictions and reweights their loss to emphasize under-represented tail-author distinctions.Hard samples are incorrectly clustered pairs, while simple samples are correctly clustered pairs.
  • Harmony Contrastive Learning: HCL uses LHCL = α × BCESimple + (1 − α) × BCEHard with α < 0.5, empirically set to 0.2, so hard samples dominate training.

Graph-Refined Distance Matrix (GRDM)

GRDM refines graph topology and adjusts pairwise similarities using node-degree confidence, complementing HCL to reduce the over-merging of tail authors into prolific clusters.

  • Motivation: Vanilla HAC can still over-merge tail authors because prolific authors dominate graph connections and GNN message passing homogenizes neighboring embeddings.
  • Topology Refinement: GRDM first refines the heterogeneous graph using HCL-derived embedding similarities, removing low-semantic-similarity noisy edges and adding high-similarity tail-author edges.
  • Degree-Aware Similarity Adjustment: Node degrees in the refined graph are converted into embedding-confidence scores ranging from 1 to 2, with higher degrees indicating higher confidence.
  • Degree-Aware Similarity Adjustment: GRDM amplifies similarity for connected pairs with similar high confidence and reduces it for connected pairs with mismatched confidence.
  • Distance Construction: Unconnected pairs retain their original similarity, preserving the global semantic structure before conversion to the final distance matrix for HAC.
  • Role in GRAND-HC: Together, HCL addresses prolific bias during representation learning, while GRDM addresses residual bias during clustering through topology and distance refinement.

Cluster Number Estimation with Paper Compression Module (PCM)

The Paper Compression Module estimates the number of author clusters for long and variable paper sequences by compressing paper embeddings before sequence modeling, after which HAC uses the estimate and GRDM distances.

  • Motivation: Reliable cluster-number estimation is necessary for end-to-end SND, but existing approaches become unreliable or difficult to scale as paper sequences grow.
  • Existing Limitations: Heuristic methods iteratively search for K and may be computationally inefficient, while DBSCAN-based methods depend sensitively on hand-crafted distance thresholds.
  • PCM Architecture: PCM decouples long-sequence modeling by first compressing extensive paper embeddings into a compact fixed-length sequence, then applying bidirectional sequence modeling.
  • Cross-Attention Compression: Learnable fixed-length queries use cross-attention with paper embeddings as keys and values to extract information relevant to cluster-number estimation.
  • Sequence and Structural Features: A Bi-LSTM summarizes forward and backward contextual information, while subgraph count and paper count provide structural auxiliary inputs to the decoder.
  • Final Clustering: HAC with average linkage receives PCM's predicted cluster size K and GRDM's distance matrix D to output final author clusters.

Algorithmic Workflow of GRAND-HC

GRAND-HC constructs heterogeneous paper graphs and iteratively refines embeddings, graph structure, distances, and cluster-size estimates before producing author clusters with HAC.

  • The final output is a set of K author clusters produced after cluster-size estimation and hierarchical agglomerative clustering.
  • GRAND-HC constructs a paper relational heterogeneous graph and initializes features from BERT embeddings, subgraph labels, node degree, and publication year.
  • The iterative workflow computes paper embeddings, updates network parameters, and optimizes graph structure.
  • The framework computes node confidence scores and a refined distance matrix before estimating cluster size over iterative updates.
  • Paper embedding generation uses O(16 · d + N · D · d), while harmony contrastive learning costs O(N^2 · d + N^2 · log(N) + TN^3 · d).Cluster-size estimation costs O(S · N · d + S · (d · h + h^2)).

Experiment Setup

Experiments evaluate GRAND-HC on AMiner-v2 and WhoisWho-v1 using macro metrics and unified fixed- or unknown-cluster-size protocols, with comparisons, ablations, and efficiency analyses.

  • Datasets: AMiner-v2 provides 400 training, 100 validation, and 100 testing names, while WhoisWho-v1 provides 221, 50, and 50 respectively.WhoisWho-v1 reaches 5682 maximum papers per name versus an average of 911.1.
  • Metrics: Evaluation uses pairwise precision, recall, and F1 for each name, then reports macro precision, recall, and F1 as final metrics.
  • Protocols and baselines: The study compares feature-based, graph-based, and LLM-based methods under unified fixed-cluster-size and unknown-cluster-number protocols.
  • Results: GRAND-HC outperforms most current methods in macro F1 on both AMiner-v2 and WhoisWho-v1.
  • Results: Under unknown cluster numbers, GRAND-HC achieves the lowest RMLSE of 0.18, negligible performance decline, and the fastest inference speed.It adds the smallest extra delay among the retained valid methods.
  • Ablations: Removing or replacing components in ablations changes performance, with HCL raising F1 from 56.41 to 78.35 and GRDM further reaching 81.25.
  • Limitations: Some WhoisWho-v1 names remain difficult because inaccurate initial graph construction requires iterative modification that increases computational complexity and reduces efficiency.The analyzed difficult names averaged 14.78 F1 score(%).

Ablation of Paper Relational Heterogeneous Graph Construction

The relational-graph ablation evaluates co-author, co-organization, and co-venue edges, finding that combining multiple relations yields the strongest reported performance.

  • Graph construction: The study incrementally adds co-author, co-organization, and co-venue edges to heterogeneous paper graphs built from AMiner and WhoisWho training data.
  • Results: The A1+O+V graph achieves F1 scores of 47.57 on AMiner and 37.16 on WhoisWho, making it the best-performing relational graph in both datasets.
  • Edge contributions: Adding co-organization information substantially improves A+O over A alone, whereas co-venue contributes minimally to disambiguation performance.The paper suggests co-venue may primarily convey semantic information such as shared research directions.
  • Graph quality: Final training performance strongly correlates with initial heterogeneous-graph precision, while GRAND-HC generalizes across many initial graph types.The paper attributes this to the reliability of connecting edges and graph-based distance refinement.

Ablation of Paper Embedding Generation

Embedding and cluster-size-estimation ablations examine harmony loss weighting, graph features, and query design, showing benefits from balanced representations and compressed sequences.

  • Harmony contrastive learning: Increasing α reduces performance because emphasizing easy pairs from prolific authors hinders feature learning for authors with fewer papers.Moderating simple-sample loss contributions produces more comprehensive paper representations.
  • Harmony contrastive learning: With α = 0.5, harmony contrastive learning becomes the traditional contrastive loss, while the visualization shows greater separation among smaller authors under the moderated weighting.
  • Graph information: Ablations evaluate subgraph labels and degree as graph-information inputs alongside direct clustering of OAG-BERT-V2 embeddings and initial graph connected components.
  • Cluster-size estimation: GRAND-HC’s cluster-size estimator is compared with representative methods using root mean squared logarithmic error on AMiner-v2.
  • Cluster-size estimation: GRAND-HC’s RMLSE is far lower than those of the other evaluated cluster-size estimation algorithms.
  • PCM analysis: Query sequences have an optimal length around 25; longer sequences add repetitive information, while shorter sequences may omit useful data.

Ablation of Graph-refined Distance Matrix

The GRDM ablation separates graph-structure optimization from graph-based distance refinement, showing that topology quality and degree-aware adjustment determine clustering robustness. Visualization and parameter studies further examine the refined distances and graph-optimization settings.

  • GRDM design: GRDM comprises graph structure optimization followed by graph-based distance matrix refinement.The method refines graph topology using embedding similarity, then adjusts pairwise distances using node degree and edge connectivity.
  • Ablation results: GHAC-o and GHAC-r depend strongly on topology accuracy, whereas GRAND-HC-o is more robust through degree-adaptive tuning.GRAND-HC-r improves more when the graph structure is accurate.
  • Parameter analysis: Increasing γ significantly improves F1, while ψ has little effect across the tested ranges.The tested ranges are ψ = 0-0.5 and γ = 0.5-0.95.
  • Iteration analysis: A single graph-optimization iteration is selected because additional iterations provide lower gains and excessive iterations slightly decrease F1.The complexity of k modifications for N nodes with d-dimensional features is O(N^2 · k · d).
  • Visualization: GRDM produces pairwise relationships closer to ground-truth text relations than the primitive cosine distance matrix.The visualization compares labeling, cosine-distance, optimized-topology, and refined-distance matrices.
Loading 2609.01636v1…