PaperlayerView exact revision on arXiv

Source-linked AI summary

Scientific Paper Recommendation: A Survey

Xiaomei Bai, Mengyang Wang, Ivan Lee, Zhuo Yang, Xiangjie Kong, Feng Xia

arXiv:2008.13538v1cs.IR

TL;DR

Scientific paper recommendation must help researchers navigate rapidly expanding scholarly literature and information overload. This survey reviews four recommendation-method families, their evaluation metrics, and open issues, concluding that content-based and hybrid methods are most often used.

  • Problem

    Rapidly increasing scholarly literature creates information overload, motivating systems that help researchers find relevant papers beyond keyword-based search.

  • Method

    The survey classifies content-based, collaborative-filtering, graph-based, and hybrid methods, then reviews evaluation metrics and open challenges.

  • Results

    Content-based and hybrid methods are the most often used techniques in the surveyed paper-recommender systems.

  • Takeaways & Limitations

    The survey provides a structured overview of recommendation rationales, advantages, disadvantages, applications, metrics, and challenges in scholarly paper recommendation.

Abstract

from arXiv · show

Globally, recommendation services have become important due to the fact that they support e-commerce applications and different research communities. Recommender systems have a large number of applications in many fields including economic, education, and scientific research. Different empirical studies have shown that recommender systems are more effective and reliable than keyword-based search engines for extracting useful knowledge from massive amounts of data. The problem of recommending similar scientific articles in scientific community is called scientific paper recommendation. Scientific paper recommendation aims to recommend new articles or classical articles that match researchers' interests. It has become an attractive area of study since the number of scholarly papers increases exponentially. In this survey, we first introduce the importance and advantages of paper recommender systems. Second, we review the recommendation algorithms and methods, such as Content-Based methods, Collaborative Filtering methods, Graph-Based methods and Hybrid methods. Then, we introduce the evaluation methods of different recommender systems. Finally, we summarize open issues in the paper recommender systems, including cold start, sparsity, scalability, privacy, serendipity and unified scholarly data standards. The purpose of this survey is to provide comprehensive reviews on scholarly paper recommendation.

I. INTRODUCTION

Scientific paper recommendation addresses information overload by matching researchers with relevant literature. This survey organizes paper-recommender methods, evaluation metrics, and open challenges, with content-based filtering using profiles and paper similarity.

  • Scientific paper recommendation ranks and recommends papers associated with a researcher’s interests or research focus.
  • Recommender systems personalize literature discovery more effectively than keyword search for massive collections.Keyword search can return identical, overly large result sets for researchers with different interests.
  • The survey classifies paper-recommendation techniques into content-based, collaborative-filtering, graph-based, and hybrid methods.It also discusses additional techniques such as latent factor and topic regression matrix-factorization models.
  • The survey analyzes evaluation metrics and summarizes challenges including cold start, sparsity, scalability, privacy, serendipity, and unified data standards.
  • A. CONTENT-BASED FILTERING (CBF): Content-based filtering builds researcher and paper representations, computes their keyword similarity, and ranks highly similar candidate papers.TF-IDF, keyphrase extraction, language, topic, and concept-based models are among the representation approaches discussed.

B. COLLABORATIVE FILTERING (CF)

Collaborative filtering recommends papers from patterns in users’ ratings, opinions, and relationships rather than paper content. Its user- and item-based variants rely on similarity and prediction, but face cold-start and relevance limitations.

  • Collaborative filtering recommends papers by using opinions or ratings from other users, even without content-based features.Ratings may come from reference-management websites or user questionnaires.
  • A user-item matrix represents ratings, enabling systems to calculate user similarity and identify neighbour users for recommendation.Matrix factorization can predict ratings for candidate papers.
  • User-based approach: User-based collaborative filtering finds neighbouring users and ranks items according to their interests.User similarity is calculated from ratings on common items, with social relations sometimes added to identify neighbours.
  • Item-based approach: Item-based collaborative filtering computes item similarity and generates predictions from the most similar items associated with the target user.Similarity may use cosine or thematic similarity.
  • Collaborative filtering can recommend papers popular among a target user’s connections, but recommendations may not match the user’s current research.
  • Cold start prevents recommending unrated new items and leaves new users without enough history to identify a similar neighbourhood.Graph-based and hybrid methods are presented as techniques intended to address these problems.

C. GRAPH-BASED METHOD (GB)

Graph-based paper recommenders represent researchers and papers as nodes connected by interest, authorship, citation, or social relationships, then rank relevant papers through graph search and related algorithms.

  • Graph Construction: Graph-based systems construct heterogeneous graphs whose nodes represent researchers and papers, with edges encoding interests and relationships between researchers or papers.Some models additionally include researcher–researcher and paper–paper edges.
  • Graph Construction: Researcher behavior data, such as publications or references, can be converted into a two-part graph to calculate relevance between unconnected researchers and papers.In the illustrated model, researcher–paper links reflect published papers and support recommendation as a graph-search task.
  • Graph Construction: Citation graphs represent papers as nodes and citation relationships as edges, enabling recommendations based on shared references or citation links.Citation networks can also be combined with content, co-author, or key-term relationships.
  • Recommendation Generation: Graph-based recommenders use graph structure rather than paper content or researcher profiles to identify relevant papers.Random walk and random walk with restart iteratively produce probability distributions used to rank graph vertices.
  • Recommendation Generation: PaperRank extends PageRank to evaluate scientific papers through indirect citation relationships and express individual paper importance as a ranking score.Its formulation uses reference counts, a damping coefficient, and an indicator for whether one paper cites another.

D. HYBRID METHOD (HM)

Hybrid methods combine content-based, collaborative-filtering, graph-based, and other techniques to improve paper recommendation by using complementary information and mitigating individual methods’ shortcomings.

  • Overview: Hybrid paper recommenders combine two or more techniques and information sources to improve personalized recommendation results.A surveyed example combines content-based and collaborative-filtering methods.
  • Content-based + Collaborative Filtering: Content-based and collaborative filtering can be combined sequentially or in parallel to use paper content, user profiles, citation similarity, and recommendation scores.One three-stage process builds a TF-based profile, selects papers by TF-IDF cosine similarity, then uses citation correlations to identify neighboring papers.
  • Content-based + Graph-based: Content-based and graph-based methods complement one another by combining researcher profiles from paper content with candidate discovery from citation or bipartite graph structure.This combination is described as performing better than classic recommendation methods.
  • Weighted Combination: Hybrid systems can weight content-based and graph-based result lists to produce a combined recommendation score.The weights are w and (1 −w), and the combination is reported to address over-specialization and new-item problems.
  • Challenges and Performance: The main challenge for hybrid methods is effectively combining techniques despite their varied combinations and information sources.The survey also reports that special hybrid methods perform better than baseline methods.

E. OTHERS

Other surveyed techniques address representation and scalability issues through modified latent-factor models, hashing-based citation representations, and semantic modeling of candidate papers.

  • Additional Techniques: Modified latent-factor models can represent users and papers using interaction data, paper content, attributes, and social-network information.A modified topic model represents users and papers before matrix factorization predicts recommendations.
  • Sparsity: Scientific paper recommendation faces sparse matrices because researchers are much fewer than papers, leaving many empty user-item or citation-matrix elements.This sparsity arises when citation or user-item matrices represent relationships across the full paper collection.
  • Sparsity: Non-sparse matrix representations and locality-sensitive hashing are used to represent citation networks and address sparse citation matrices.The cited matrix examples preserve the same citation relationships while using hash functions to determine similarity.
  • Semantic Modeling: Other approaches use preprocessing with content-based filtering and LSTM models to learn semantic representations of candidate papers and rank highly similar results.The described process selects top-N papers using content and semantic similarity.

F. COMPARISONS OF COMMON TECHNIQUES

The survey compares content-based, collaborative-filtering, and graph-based techniques and presents hybrid methods as combinations intended to offset their respective disadvantages.

  • Common Techniques: Content-based filtering, collaborative filtering, and graph-based methods each have advantages and disadvantages that can complement one another.The survey organizes these comparisons in Table 4.
  • Common Techniques: Collaborative filtering can address recommendation-quality concerns but still has cold-start and other disadvantages.Hybrid methods combine collaborative filtering with content-based or graph-based methods to improve efficiency and avoid disadvantages.

III. EVALUATION METHODS

Evaluation methods determine whether scientific paper recommendation techniques work effectively and use metrics suited to different recommendation approaches. The survey reviews ranking, accuracy, coverage, and list-based evaluation measures.

  • Precision and Recall are the most frequently used evaluation methods in the reviewed paper recommender systems.
  • Precision measures the fraction of recommended papers that are relevant to researchers.A larger value indicates more accurate recommendations; P@N is a modified version for the top N results.
  • Recall measures the fraction of all relevant papers that appears in the recommendation result list.Recall@m counts relevant papers within the top m positions, and larger values indicate better ranking of relevant papers toward the top.
  • F-measure combines Precision and Recall as a weighted harmonic average because increasing list size can raise Recall while reducing Precision.A high F value indicates a more effective paper recommendation system.
  • NDCG, MAP, and MRR evaluate the quality or ranking of sorted recommended paper lists from complementary perspectives.NDCG emphasizes gain by rank position, MAP averages users’ average precision, and MRR focuses on relevant-paper ranks.
  • RMSE and MAE evaluate rating-prediction accuracy, while UCOV measures the proportion of users receiving relevant recommendations.Lower RMSE and MAE indicate better prediction, whereas UCOV reflects usefulness across users.

IV. OPEN ISSUES AND CHALLENGES

Existing scientific paper recommender systems provide useful papers but still face several unresolved problems and challenges. The survey identifies six open issues requiring further improvement.

  • The survey identifies Cold Start, Sparsity, Scalability, Privacy, Serendipity, and Unified data standards as open issues.

A. COLD START

Cold start concerns new papers and new users in scientific paper recommender systems. Collaborative filtering struggles with both, while content-based methods address new papers but depend on sufficient user history.

  • Collaborative filtering faces cold-start challenges for both new users and newly published papers.New users may lack ratings or similar neighbors, while new papers have few readers and ratings for accurate recommendation.
  • Content-based recommendation can overcome the new-paper problem by analyzing paper content and comparing it with user profiles.
  • Content-based recommendation becomes unreliable when a researcher’s historical records provide insufficient information to build a useful profile.

B. SPARSITY

Scientific paper recommender systems often operate with fewer users than papers, producing sparse rating matrices. This sparsity makes user similarity and collaborative filtering difficult.

  • The number of users is often less than the number of papers, contrary to an assumption used by many recommender systems.
  • Collaborative filtering encounters sparse rating matrices because users rate only a few papers and many papers receive few ratings.
  • Sparse ratings make it difficult to identify similar user neighbors, creating a major disadvantage for collaborative filtering systems.

C. SCALABILITY

Scalability concerns whether recommender systems can work effectively with numerous users and products, especially as digital-library datasets grow and change dynamically.

  • C. SCALABILITY: Digital-library datasets are large and dynamic because papers and users are added every day.These changing datasets make it challenging for paper recommender systems to operate effectively.

D. PRIVACY

Privacy is a central challenge because personalized paper recommenders rely on user information that may be sensitive. Scholarly data also vary across platforms, while serendipitous recommendations may benefit researchers beyond their established interests.

  • D. PRIVACY: Personalized paper recommenders collect user information to address information overload, but that information may be sensitive and users may dislike excessive data collection.Secure recommender systems have been proposed to protect private information while using limited data carefully.
  • D. PRIVACY: Serendipitous recommendations can help junior researchers broaden their research range and senior researchers discover knowledge from other areas.Traditional systems generally recommend papers relevant to users’ existing interests or research.
  • D. PRIVACY: Scholarly data from platforms and datasets differ in their available information, such as citation relationships, complicating paper-recommender construction.DBLP lacks citation relationships, whereas APS provides citation relationships between papers.
  • D. PRIVACY: The survey identifies privacy, serendipity, and unified scholarly data standards among the open challenges for scientific paper recommendation.It also reviews cold start, sparsity, and scalability as future challenges.
Loading 2008.13538v1…