PaperlayerView exact revision on arXiv

Source-linked AI summary

Task-Guided and Path-Augmented Heterogeneous Network Embedding for Author Identification

Ting Chen, Yizhou Sun

arXiv:1612.02814v2cs.LGcs.AIcs.IRstat.ML

TL;DR

The paper addresses author identification from anonymized papers in double-blind review, where existing embeddings are often task-independent and limited to homogeneous networks. It proposes task-guided, path-augmented heterogeneous network embedding with joint objectives and task-based meta-path selection, achieving significantly better accuracy than existing methods. The approach remains limited by sparse paper information and incomplete text, while human experts may be more accurate in familiar domains.

  • Problem

    Author identification must rank potential authors from anonymized papers, but existing embeddings are usually task-independent and designed for homogeneous networks.

  • Method

    The model jointly trains shared node embeddings with task-specific and network-general objectives while selecting meta paths according to author-identification performance.

  • Results

    The combined model achieves significantly better accuracy at identifying true authors than existing methods, while experiments support task guidance and meta-path usefulness.

  • Takeaways & Limitations

    Task guidance can make network embeddings more suitable for author identification, and meta-path selection can exploit heterogeneous network information for this task.

  • Takeaways & Limitations

    The approach operates despite severe information scarcity, and the authors suggest improvement through more complete paper information and advanced text understanding; human experts may remain more accurate in familiar domains.

Abstract

from arXiv · show

In this paper, we study the problem of author identification under double-blind review setting, which is to identify potential authors given information of an anonymized paper. Different from existing approaches that rely heavily on feature engineering, we propose to use network embedding approach to address the problem, which can automatically represent nodes into lower dimensional feature vectors. However, there are two major limitations in recent studies on network embedding: (1) they are usually general-purpose embedding methods, which are independent of the specific tasks; and (2) most of these approaches can only deal with homogeneous networks, where the heterogeneity of the network is ignored. Hence, challenges faced here are two folds: (1) how to embed the network under the guidance of the author identification task, and (2) how to select the best type of information due to the heterogeneity of the network. To address the challenges, we propose a task-guided and path-augmented heterogeneous network embedding model. In our model, nodes are first embedded as vectors in latent feature space. Embeddings are then shared and jointly trained according to task-specific and network-general objectives. We extend the existing unsupervised network embedding to incorporate meta paths in heterogeneous networks, and select paths according to the specific task. The guidance from author identification task for network embedding is provided both explicitly in joint training and implicitly during meta path selection. Our experiments demonstrate that by using path-augmented network embedding with task guidance, our model can obtain significantly better accuracy at identifying the true authors comparing to existing methods.

1. INTRODUCTION

The paper frames author identification in double-blind review as a heterogeneous-network problem and proposes task-guided, path-augmented network embedding to address limitations of feature engineering and general-purpose homogeneous embeddings.

  • Motivation: Double-blind review anonymizes authors while exposing paper content and references, enabling author identification from these observable attributes.The issue raises questions about the effectiveness of double-blind review.
  • Motivation: Feature engineering is time-consuming, problem specific, non-transferable, and potentially too simple for complicated networked data.Network embedding instead automatically learns vector representations for network nodes.
  • Challenges: Existing embeddings are generally task-independent and designed for homogeneous networks, creating challenges for task guidance and heterogeneous information selection.Heterogeneous-network connections can carry richer, path-dependent semantics.
  • Approach: The proposed model jointly trains shared node embeddings with author-identification and network-general objectives.This combines task-specific author scoring with neighbor prediction that preserves network structure and proximities.
  • Approach: Meta-path selection provides implicit task guidance, while joint learning provides explicit guidance for heterogeneous network embedding.The framework is intended for author identification and other tasks.
  • Findings: The learning algorithm is efficient and parallelizable, and experiments report much better results than existing feature-based methods.The experiments also demonstrate the effectiveness of task guidance and meta-path selection.

2. PRELIMINARIES

The preliminaries define heterogeneous bibliographic networks, meta paths, node embeddings, and author identification as ranking potential authors from an anonymized paper’s observed neighbors.

  • 2.1 Heterogeneous Networks: A heterogeneous network contains multiple node types and/or link types, represented with node- and link-type mapping functions.The bibliographic network includes paper, author, keyword, venue, and year nodes alongside typed links.
  • 2.1 Heterogeneous Networks: The bibliographic network is centered on papers, whose neighboring nodes represent information such as authors, keywords, venues, and years.Figure 2 schematizes node types and link types in this network.
  • 2.1 Heterogeneous Networks: A meta path is a schema-level path expressing a compositional relation between node types through a sequence of link types.Different paths between the same node type can encode different semantic meanings.
  • 2.2 Embedding Representation of Nodes: Network embedding maps nodes into a common latent feature space, enabling direct similarity or distance calculations from their vectors.The embedding table U has one D-dimensional vector for each of N nodes.
  • 2.3 Author Identification Problem: Author identification ranks potential authors for each anonymized paper using its observed neighboring nodes, including keywords, references, venue, and year.The target is for top-ranked authors to belong to the paper’s true-author set.

3. PROPOSED MODEL

The proposed model combines task-specific author identification with path-augmented heterogeneous network embedding. Shared node embeddings are jointly trained and meta paths are selected according to the author identification task.

  • 3. PROPOSED MODEL: The model has task-specific author identification and path-augmented general network embedding components combined into one framework.Meta paths are selected according to the author identification task.
  • 3.1 Task-Specific Embedding for Author Identification: The task-specific model maps nodes into latent space, aggregates observed neighbor embeddings into a paper representation, and scores potential authors.The paper representation is built through two aggregation stages: averaging embeddings by node type, then combining node-type representations.
  • 3.2 Path-Augmented General Heterogeneous Network Embedding: The path-augmented model uses meta-path-induced adjacency matrices and conditions neighbor prediction on both the source node and path type.Candidate paths can encode different semantics, such as shared topics or affiliations.
  • 3.2 Path-Augmented General Heterogeneous Network Embedding: Negative sampling approximates neighbor-distribution evaluation when destination node sets are very large.Negative nodes are sampled from a predefined noise distribution, with path-specific bias terms adjusting densities.
  • 3.3.2 Meta Path Selection - An Implicit Guidance: Meta paths are selected by maximizing author-identification generalization performance on validation data.This provides implicit task guidance, complementing explicit guidance from the joint objective.
  • 3.3.1 Joint Objective - An Explicit Guidance: The two objectives are jointly optimized with shared embeddings using a weighted linear combination and regularization.The trade-off factor ranges from 0 to 1, with the endpoints selecting network-general or supervised embedding alone.

4. EXPERIMENTS

Experiments evaluate the proposed model on the AMiner citation network using temporally split papers, heterogeneous-network baselines, ranking metrics, and meta-path analyses. Results show that task-guided path selection improves author identification, with especially strong benefits for papers involving less frequent authors.

  • Data: The experiments use AMiner papers published before 2014 for training and papers from 2014–2015 for testing.A heterogeneous bibliographic network is extracted from paper titles, authors, references, venues, and keywords.
  • Baselines and settings: The evaluation compares supervised feature-based methods with task-specific, network-general, pre-trained, and combined embedding models.The combined model integrates task-specific embedding with meta-path selection-based network-general embedding.
  • Evaluation metrics: MAP@k measures the accuracy of top-ranked authors, while Recall@k measures retrieval of true authors among the top k results.Average precision is computed for each test paper and then averaged across papers for MAP@k.
  • Meta-path selection: Author-identification performance rises when the first several meta-paths are added, then declines as additional paths introduce less relevant or noisy information.The selected paths for joint learning are A-P-P, A-P-W, and P-A.
  • Meta-path selection: Using all length-2 paths worsens results, whereas greedy-selected paths improve embedding models, demonstrating that task-related path selection matters.The comparison reports performance under different network paths using MAP@3 / Recall@3.
  • Performance comparison: The proposed method significantly outperforms feature-based and embedding baselines, especially on papers with less frequent authors.The authors conjecture that pre-training or joint unsupervised embedding alleviates over-fitting in task-specific embedding.

5. DISCUSSION

The discussion reports reasonable top-rank author identification despite sparse paper information, while performance varies with author publication history and double-blind review may protect junior researchers.

  • Sparse paper information still permits reasonable top-rank identification among one million candidate authors.The paper notes a median of one reference per paper and uses only keywords.
  • Authors with more past publications are easier to predict, whereas authors with few publication records are substantially harder.The paper interprets this pattern as highly visible authors being easier to detect.
  • Figure 8 examines the combining factor between network-specific and network-general objectives and speedup versus thread count.
  • The authors suggest double-blind review still helps protect junior researchers.

6. RELATED WORK

Related work covers heterogeneous-network mining, network embedding, and author identification, contrasting prior feature engineering with this paper’s automatic feature learning.

  • Heterogeneous-network research uses meta paths to represent networks containing multiple node or link types.Prior work spans bibliographic, recommendation, and social networks.
  • Many heterogeneous-network mining methods rely on feature engineering, whereas this paper adopts embedding methods for automatic feature learning.
  • Network embedding methods are associated with scalable representations and include approaches for both homogeneous and heterogeneous networks.
  • Prior author-identification work differs from this setting because existing authors are unknown here and paper references are used as information.

7. CONCLUSION AND FUTURE WORK

The conclusion frames author identification as ranking authors for anonymized papers in heterogeneous networks and presents joint task-guided, path-augmented embedding as the proposed solution.

  • The model jointly trains shared node embeddings with task-specific author identification and network-general objectives.
  • Meta paths are incorporated into heterogeneous network embedding and selected according to the author identification task.
  • Experiments show meta paths are useful and combining both objectives significantly improves true-author identification accuracy over existing methods.
  • Future work includes author-set prediction that models author interactions and deeper text analysis using full papers.

A. FEATURE ENGINEERING FOR TRADITIONAL SUPERVISED MODELS

The traditional supervised baselines use author features and paper–author paired features to rank authors for a paper, covering publication, reference, vocabulary, venue, and temporal signals.

  • The traditional supervised models use both author features and paper–author paired features for author ranking.
  • Author-level features include total papers, recent publications, distinct venues, and distinct publication years.
  • Paper–author paired features measure prior citations and references involving the author, using counts and ratios.
  • Author-level reference features measure the number and ratio of references written by the author.
  • Other author features measure shared and unique words, their ratios, and prior attendance at the venue.

B. DERIVATION OF TASK-SPECIFIC EMBEDDING FOR AUTHOR IDENTIFICATION

The task-specific embedding model computes parameter gradients and learns from paper information and true author sets through iterative triple sampling and parameter updates.

  • The model’s parameter gradients are calculated for task-specific embedding.
  • Algorithm 2 takes paper information X and the true author set A as input and outputs parameters U and w.
  • Training iterates until convergence, sampling triples (p, a, a′) and updating U and w according to Eqs. 10 and 11.

C. DERIVATION OF PATH-AUGMENTED GENERAL HETEROGENEOUS NETWORK EMBEDDING

The path-augmented general heterogeneous network embedding model computes parameter gradients and learns from path-derived adjacency matrices through iterative positive and negative node sampling.

  • The model’s parameter gradients are calculated for path-augmented general heterogeneous network embedding.
  • Algorithm 3 takes path adjacency matrices {M} derived from heterogeneous network G as input and outputs parameters U and b.
  • Training iterates until convergence, sampling triples (r, i, j), negative nodes {j′}, and updating U and b according to Eqs. 12 and 13.
Loading 1612.02814v2…