PaperlayerView exact revision on arXiv

Source-linked AI summary

Towards Automatic Evolution Tree Generation from Citation Graphs

Zexing Zhao, Yuntong Hu, Liang Zhao

arXiv:2609.09561v1cs.CL

TL;DR

Existing methods struggle to represent scientific evolution because taxonomy induction is often leaf-bound and time-agnostic, making transitional papers difficult to place. EvoTree separates conceptual backbone learning from temporal refinement, re-attaches marginal papers to internal nodes under temporal constraints, and labels concepts with an LLM without changing topology. On an annotated benchmark spanning 11 AI subfields, it reports the strongest overall structural and concept-alignment results and non-trivial marginal-paper detection.

  • Problem

    Existing taxonomy-induction methods are largely leaf-bound and time-agnostic, while transitional papers require temporally consistent placement in scientific evolution trees.

  • Method

    EvoTree learns a graph-aware distributional taxonomy backbone, temporally refines it with constrained internal-node re-attachment, calibrates with few-shot metric learning, and uses an LLM only for labeling.

  • Results

    EvoTree attains the best NMI, citation-direction accuracy, and annotated concept purity, and is the only method with non-trivial marginal-paper detection on the annotated benchmark.

  • Takeaways & Limitations

    The framework supports evolution trees that preserve conceptual structure while representing temporal development and transitional papers.

  • Takeaways & Limitations

    Evaluation uses AI survey data, and transfer to disciplines with substantially different citation conventions and paradigm cadences has not been verified.

Abstract

from arXiv · show

Surveys remain the primary way researchers grasp the lineage of methods within an AI subfield, but they scale poorly against the current rate of publication. Existing taxonomy-induction methods are largely leaf-bound and time-agnostic; they tend to force transitional papers into mature leaves and can create topological inversions between ancestors and descendants. We propose EvoTree, a staged framework that decouples conceptual backbone learning from temporal refinement: a graph-aware encoder with distribution-based hierarchical clustering yields a stable taxonomy backbone; temporal fine-tuning then re-attaches marginal papers to internal nodes under monotonic-path constraints; a final LLM pass labels concepts without altering the topology. We release the first annotated benchmark for this task across 11 AI subfields. EvoTree attains the highest NMI and citation-direction accuracy among all baselines and the best concept purity on the annotated benchmark, and is the only method with non-trivial marginal-paper detection on the annotated set.

1 Introduction

The paper frames automatic evolution-tree generation as a response to rapidly expanding literature and the limitations of manual, time-agnostic taxonomy induction. EvoTree separates stable conceptual organization from temporal refinement and handles marginal papers explicitly.

  • Rapid literature growth makes it difficult for researchers to understand historical evolution, conceptual lineage, and paradigm transitions efficiently.
  • Evolution trees expose conceptual inheritance, branching, and progressive refinement, but their construction remains largely manual and expertise-intensive.
  • The task asks whether scientific evolution trees can be generated automatically from citation graphs.
  • Citation graphs must be transformed into hierarchical, temporal, and semantically structured representations while preserving citation dependencies and scientific progression.
  • Marginal or transitional papers are difficult to place in conventional taxonomies without creating chronological inconsistencies.
  • EvoTree first learns a stable time-agnostic taxonomy, then performs temporally constrained refinement with internal-node re-attachment for marginal papers.The framework combines distribution-based clustering, temporal admissibility, and structural legality constraints.

2 Related Work

Prior research organizes concepts through taxonomy induction, models temporal topic evolution, and analyzes citation trends, but does not directly construct temporally constrained scientific evolution trees.

  • Automatic taxonomy induction mainly expands or completes existing hierarchies rather than modeling scientific evolution directly.
  • Temporal topic models capture topic drift and time-varying word distributions across document collections.
  • Citation and co-citation analyses primarily characterize global scientific trends, while recent taxonomy-like summaries optimize textual synthesis rather than explicit structural evolution.
  • Existing methods lack temporal constraints over evolutionary paths and mechanisms for handling marginal papers or constructing survey-style evolution structures.

3 Problem Formulation

The paper formalizes evolution-tree generation over citation graphs by extending conceptual taxonomies with temporally admissible concept relations and flexible paper attachments. Marginal papers are identified through weak leaf compatibility and may attach to internal concepts.

  • A citation graph contains referenced papers as nodes with text and publication years, while directed edges indicate citation relations.
  • The target survey is excluded from the citation graph used for the main evaluation.
  • A taxonomy tree organizes papers by conceptual inclusion, with paper nodes attached as terminal leaves and edges encoding conceptual refinement without temporal semantics.
  • An evolutionary tree extends the taxonomy with a rooted concept tree, temporally admissible development relations, and paper attachments that may target internal concept nodes.
  • Temporal statistics over publication years support structural constraints and path-level evaluation of concept development.
  • A paper is marginal when its best compatibility with all leaf concepts falls below threshold η, after which it may attach to an internal node rather than being discarded.

4 Methodology

EvoTree uses a shared graph-aware encoder and distributional tree builder across staged taxonomy learning, temporal refinement, few-shot calibration, and concept labeling. The stages preserve a conceptual backbone while enforcing temporal legality and accommodating marginal papers.

  • Overview: EvoTree encodes citation-graph nodes with paper text and timestamps, then uses a graph-aware encoder and distributional tree builder as shared backbone modules.
  • Distributional Tree Builder: Distributional clustering forms base clusters and hierarchically agglomerates Gaussian summaries using a closed-form 2-Wasserstein distance.
  • Distributional Tree Builder: Re-estimating parent Gaussians after merges reduces level-wise drift, while the same builder is reused across Stages I and II with different supervision and temporal admissibility.
  • Stage I: Taxonomy Backbone: Stage I learns a time-agnostic taxonomy from section-derived weak supervision, citation consistency, and cluster geometry, producing the structural prior for Stage II.
  • Stage II: Evolutionary Refinement: Stage II alternates structure and representation updates while adding temporal admissibility, citation directionality, marginal-paper re-attachment, and constraints preserving temporal consistency, backbone branches, and tree legality.The method uses lightweight adapters while keeping the heavy SPECTER2 and GNN components frozen.
  • Marginal-Paper Handling: Marginal papers are detected through low best-leaf compatibility and re-attached to the Gaussian-supported concept among all nodes, allowing broader internal attachments.
  • Calibration and Labeling: Stage III applies few-shot metric-learning calibration to reference trees, and a final LLM pass verbalizes concept labels without changing topology or paper attachments.The calibration updates only a lightweight adaptation layer and shapes embedding geometry rather than predicting discrete labels.

5 Experiments

EvoTree is evaluated on semantic clustering, evolution structure, annotated few-shot alignment, training-stage contributions, citation-graph reliance, and direct generation quality. It achieves strong overall results while exposing capabilities that taxonomy and LLM baselines lack, especially marginal-paper detection and explicit evolutionary structure.

  • Datasets and metrics: The evaluation uses 411 survey-reference ego-graphs plus 11 manually annotated evolutionary trees spanning major AI subfields.The annotated benchmark includes concept membership, marginal-paper, evolution-edge, and node-depth labels.
  • Datasets and metrics: Metrics cover leaf and path semantic purity, NMI, citation-direction accuracy, path monotonicity, concept purity, descendant ordering, and marginal-paper AUROC.The few-shot metrics are FS-CP, FS-EDA, and FS-Mar.
  • Semantic clustering quality: Context-Aware leads LP at 0.713 and PP at 0.742, while EvoTree achieves the highest NMI at 0.526.EvoTree trades some leaf-level semantic compactness for evolution-consistent structure.
  • Evolution structure: None of the taxonomy-generation baselines produces evolution edges or marginal scores, preventing evaluation on the FS-aligned structural metrics.Hu-CiteTaxo has the highest PMR at 0.322, but its CDA is not comparable because it does not expose directional edges.
  • Few-shot benchmark: On the annotated benchmark, EvoTree achieves the highest FS-CP, modestly exceeds the strongest LLM baseline, and consistently outperforms clustering baselines.Temporal fine-tuning and tree refinement substantially improve alignment over EvoTree (Pre-train).
  • Few-shot benchmark: EvoTree matches the strongest LLM baseline on FS-EDA and uniquely provides non-trivial FS-Mar scores through explicit marginality modeling.LLM-only methods lack structure-level constraints and marginality estimates.
  • Training-stage ablation: Post-training raises FS-Mar from 0.662 to 0.750, while FS-correction recovers FS-EDA from 0.880 to 0.949 after its post-training dip.FS-CP improves by +0.104 after post-training and a further +0.118 after FS-correction; gains transfer to unseen test graphs.
  • Citation-graph ablation: Removing citation edges creates a larger validation gap, indicating that EvoTree uses the citation graph throughout optimization rather than merely propagating SPECTER2 embeddings.The MLP variant’s slightly higher PMR is attributed to temporally monotonic paths without citation constraints.

6 Conclusion

EvoTree formulates automatic scientific evolution-tree induction from citation graphs and evaluates it with an annotated benchmark spanning 11 AI subfields. Its staged design combines stable taxonomy construction, temporal refinement, marginal-paper reattachment, and topology-preserving concept labeling.

  • EvoTree formalizes scientific evolution-tree induction from citation graphs and introduces an annotated benchmark covering 11 AI subfields.
  • The framework first learns a stable taxonomy with graph-aware encoding and distribution-based hierarchical clustering, then refines it temporally.
  • Temporal refinement reattaches marginal papers to internal nodes under monotonic-path constraints, while the final LLM pass verbalizes concepts without changing topology.
  • EvoTree achieves the best NMI, citation-direction accuracy, and annotated concept purity, while uniquely recovering transitional papers with non-trivial accuracy.
  • Ablations indicate that both the citation graph and staged training framework are jointly necessary.

Limitations

The evaluation is scoped to AI survey-derived graphs and reference trees, with several limitations affecting transfer, granularity, supervision, temporal reliability, and labeling dependence.

  • The 411 ego-graphs and 11 reference trees come from AI surveys, so transfer to disciplines with different citation conventions or paradigm cadences remains unverified.
  • The few-shot dataset contains 11 curated trees and 352 papers, limiting fine-grained per-domain alignment analyses.
  • Stage I assumes survey section organization reflects a coherent conceptual taxonomy, but application- or chronology-based surveys may provide weaker supervision.
  • Publication years conflate arXiv, conference, and journal dates, adding noise to temporal admissibility and PMR and contributing to modest absolute PMR values.
  • Concept labels depend on GPT-4.1-mini; replacing it with an open-source LLM is straightforward but may change labeling style.

A.3 Analysis and Discussion

In a text-to-image diffusion attack survey, EvoTree recovers a foundation-to-application organization in which diffusion-model foundations precede downstream attack and editing work.

  • The red foundation branch temporally and conceptually precedes attack and editing work in the blue and purple branches.

B.1 Main Dataset

The main dataset combines survey-centered citation graphs with weak section-derived labels, while the few-shot benchmark supplies annotated concepts, evolution relations, marginal status, and depth for cross-domain evaluation.

  • Main Dataset: The survey-reference dataset contains 411 ego-graphs whose nodes are cited papers and whose directed edges are citations among those references.
  • Main Dataset: Each paper node includes a title, abstract, publication year, and 768-dimensional SPECTER2 embedding.
  • Main Dataset: Weak hierarchical labels are inherited from survey section paths, and graphs—not individual nodes—are split into train, validation, and test sets at 70/15/15.
  • Few-shot Benchmark: The annotated benchmark covers 11 AI subfields and records concept membership, marginal-paper status, directed evolution edges, and paper depth.
  • Few-shot Benchmark: Evaluation uses leave-one-domain-out calibration, with each held-out domain excluded from calibration, validation, threshold selection, and hyperparameter tuning.
  • Cross-domain Overlap: The largest cross-domain overlap is 3 papers, or 8.6% relative to the smaller domain, making within-set memorization unlikely while not excluding all information sharing.
  • Concept Labeling: The final LLM labeling pass uses bottom-up prompts: leaf labels are specific method phrases, while internal labels generalize over child concepts without repeating them.

D.2 Few-shot metrics on annotated graphs

The annotated-graph evaluation measures concept alignment, evolutionary direction, and marginal-paper detection. FS-CP focuses on leaf-cluster concept purity, FS-EDA uses annotated evolution edges, and FS-Mar evaluates marginality scores.

  • Evaluation metrics: The annotated evaluation distinguishes concept purity, evolutionary direction, and marginal-paper detection as separate few-shot capabilities.These metrics operate on annotated graphs rather than only citation-graph structure.
  • FS-CP — FS Concept Purity: FS-CP measures alignment between predicted leaf clusters and single-label research concepts.It is the primary structural-alignment metric in the evaluation.
  • FS-EDA — FS Evolutionary Direction Accuracy: FS-EDA evaluates whether annotated evolutionary descendants are placed deeper than their predecessors.It replaces noisy citation edges with annotated evolutionary edges representing extension or improvement.
  • FS-Mar — FS Marginal AUROC: FS-Mar uses per-paper marginality scores to evaluate detection of papers at domain boundaries.Standard clustering and LLM baselines do not produce marginality scores, so their results are reported as unavailable.

F.1 Forward computation

The forward computation builds paper representations from semantic and citation-graph information, clusters them distributionally into a taxonomy, and refines the tree with temporal and evolutionary constraints. The resulting framework also trains auxiliary objectives for taxonomy structure, evolution, and marginal-paper handling.

  • Representation learning: A pretrained scientific encoder supplies fixed semantic representations, which are processed through parallel semantic and graph branches.The branches are combined with a learnable connectivity-biased fusion gate before projection.
  • Representation learning: Isolated nodes bypass undefined neighborhood aggregation through a pure semantic residual.This preserves semantic information when a paper has no graph neighbors.
  • Distributional tree builder: Soft HDBSCAN memberships are converted into diagonal Gaussian cluster representations for distribution-based hierarchical clustering.Clusters are greedily merged using squared 2-Wasserstein distance, with parent distributions re-estimated after each merge until one root remains.
  • Distributional tree builder: The tree builder can enforce a temporal admissibility filter during merging and returns Gaussian parameters at every hierarchical node.The filter limits parent-child temporal violations to the specified tolerance.
  • Pretraining objectives: Pretraining combines section-label supervision with link prediction and contrastive ranking to preserve semantic and citation structure.The contrastive objective acts on projection representations rather than clustering embeddings, keeping clustering geometry untouched.
  • Evolution refinement: Evolution training enforces semantic coherence, forward-progressing paths, citation directionality, marginal-paper attachment, and sibling separation.The evolution loss is decomposed into semantic, temporal, and marginal terms.
  • Evolution refinement: The path constraint prevents global temporal inversions by requiring nondecreasing earliest publication years along root-to-leaf paths within a per-edge tolerance.The formulation controls cumulative violations across an entire path rather than only local parent-child pairs.
  • Evolution refinement: A taxonomy-preservation constraint requires overlap between evolutionary and taxonomy ancestor chains, using θov = 0.3.This limits temporal refinement from rewiring concepts into semantically unrelated branches.

N.1 Evaluation Setup

The evaluation compares anonymized generated hierarchies using identical inputs and standardized criteria. An LLM judge and independent human evaluators score conceptual organization, scientific evolution, transitional-paper placement, and overall quality.

  • Evaluation inputs: Both evaluation protocols receive paper titles, publication years, citation relations, and anonymized generated hierarchies.Method names, model descriptions, and metric results are withheld, and structures are randomized to reduce presentation bias.
  • Evaluation protocols: The LLM judge evaluates each structure over multiple independent runs under a fixed prompt and decoding configuration.Human evaluators independently assess structures without discussing individual cases beforehand.
  • Evaluation dimensions: Both protocols score conceptual organization, scientific evolution, transitional-paper placement, and overall explanatory quality on a 1–10 scale.Scores are averaged over runs or evaluators and then over the annotated domains, with means and standard deviations reported per method and dimension.
Loading 2609.09561v1…