Scientific Paper Recommendation: A Survey
(2008.13538v1)
Published 10 Aug 2020 in cs.IR
Abstract: 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.
This paper, "Scientific Paper Recommendation: A Survey" (Bai et al., 2020), provides a comprehensive overview of scientific paper recommender systems. It highlights the increasing need for these systems due to the exponential growth of scholarly literature and the resulting information overload faced by researchers. The survey contrasts recommender systems with traditional keyword-based search, noting that recommender systems offer a more personalized and effective approach by considering user interests, co-author relationships, and citation relationships.
The paper classifies existing scientific paper recommendation techniques into four main categories: Content-Based Filtering (CBF), Collaborative Filtering (CF), Graph-Based Method (GB), and Hybrid Methods (HM).
Content-Based Filtering (CBF) focuses on recommending items similar to those a user has liked in the past. For scientific papers, this involves three main steps:
Item Representation: Papers are represented by extracting features from their content (title, abstract, keywords, full text). Common methods include TF-IDF (Term Frequency-Inverse Document Frequency) to weight the importance of terms and keyphrase extraction models. TF-IDF calculates a score for each term based on its frequency in a specific paper and its rarity across the entire corpus.
Profile Learning: A user's profile is built based on their historical actions, such as papers they have read, cited, or published. This profile represents the user's research interests, often using the same feature vectors derived from papers. Methods include extracting keywords from past publications or using information from mind maps. The paper notes methods to differentiate profiles for junior researchers (using cited papers) and senior researchers (using cited and citing papers).
Recommendation Generation: The similarity between the user profile and candidate papers is computed, often using cosine similarity between their feature vectors. Papers with high similarity scores are ranked and recommended. Some CBF systems also explore serendipitous recommendations by considering papers from less related fields.
Advantages of CBF include its ability to recommend new items (if their content is analyzable) and provide personalized results without needing data from other users. However, it struggles with evaluating item quality (like authoritativeness or writing style) and the "new user problem" where insufficient historical data makes profile building difficult.
Collaborative Filtering (CF) recommends items based on the preferences of similar users. The core idea is that users who liked similar items in the past will likely like similar items in the future. This method typically relies on a user-item matrix representing user ratings or interactions with papers. CF is divided into two main approaches:
User-based approach: Finds users similar to the target user based on their shared item ratings. The similarity is calculated using metrics like Pearson correlation coefficient. Recommendations are then generated from items liked by these "neighbour" users.
Item-based approach: Identifies items similar to those the target user has liked. Item similarity is calculated based on the ratings given by users to pairs of items. Recommendations are based on items similar to those in the user's history with high predicted ratings.
CF's advantage is its potential for serendipitous recommendations (finding relevant items outside the user's immediate content focus) and the ability to consider item quality implicitly through user ratings. Its major drawbacks are the "cold start problem" for new users (no rating history) and new items (no ratings), and the "sparsity problem" where the user-item matrix is very sparse due to users interacting with only a small fraction of available papers.
Graph-Based Method (GB) constructs a graph where nodes represent entities like researchers and papers, and edges represent relationships (e.g., citations, co-authorship, interest). Recommendation becomes a task of searching or traversing this graph to find relevant papers.
Graph Construction: Data from digital libraries (like IEEE Xplore, CiteULike) is used to build heterogeneous graphs. This can include bipartite graphs connecting researchers and papers based on interest, citation graphs connecting papers based on citation links, or networks combining citation, co-authorship, and author-paper relationships.
Recommendation Generation: Algorithms like Random Walk with Restart (RWR) or PaperRank (an adaptation of PageRank) are applied to the graph structure to compute relevance scores between users and papers or between papers themselves. RWR simulates a random surfer moving through the graph, occasionally returning to the starting node (user). PaperRank evaluates the importance of papers based on the structure of the citation network.
GB methods can leverage information from various sources and complex relationships. However, they often do not directly consider the specific content of papers or detailed user interests beyond their interactions within the network structure.
Hybrid Method (HM) combines two or more recommendation techniques to improve performance and mitigate the disadvantages of individual methods. Common combinations include:
Content-based + Collaborative Filtering: Uses CBF to build initial profiles or recommend based on content, and CF to find similar users or items based on interaction history, addressing cold start and sparsity to some extent. Methods include combining score lists, using one method to generate candidates for the other, or running them in parallel and merging results (e.g., CBF-Separated, CF-CBF Separated, CBF-CF Parallel).
Content-based + Graph-based: CBF analyzes content for user profiles or item representations, while GB explores relationships in networks (like citation or social graphs) to find candidates or rank results.
Hybrid methods generally aim for better accuracy and overcome some limitations of single techniques by leveraging diverse information sources and algorithms. The challenge lies in effectively combining different techniques.
The paper also briefly mentions other techniques like latent factor models, hash maps for efficient sparse matrix representation, bibliographic coupling, and the application of deep learning (LSTM, CCA, SVD).
Evaluation Methods are crucial for assessing the effectiveness of paper recommender systems. The survey details commonly used metrics:
Precision: Measures the fraction of recommended papers that are relevant. Precision=TotalrecommendedpapersRelevantpapers. P@N considers precision in the top N recommendations.
Recall: Measures the fraction of all relevant papers that are successfully recommended. Recall=TotalrelevantpapersRelevantpapers. Recall@m considers recall in the top m recommendations.
F-measure: Combines Precision and Recall into a single score, providing a harmonic mean.
NDCG (Normalized Discounted Cumulative Gain): Evaluates the quality of a ranked list, giving higher scores if more relevant items appear higher in the list.
MAP (Mean Average Precision): Calculates the average precision at each point where a relevant item is retrieved, averaged over all relevant items, and then averaged across all users.
MRR (Mean Reciprocal Rank): Measures the ranking quality by taking the reciprocal of the rank of the first relevant item, averaged over all users or queries.
RMSE (Root Mean Square Error): Measures the difference between predicted ratings and actual ratings. Lower values indicate higher accuracy.
MAE (Mean Absolute Error): Similar to RMSE, it measures the average absolute difference between predicted and actual ratings. Lower values indicate higher accuracy.
UCOV (User Coverage): Measures the fraction of users for whom the system can provide recommendations.
Precision and Recall are highlighted as the most frequently used metrics in the reviewed papers.
Finally, the survey discusses several Open Issues and Challenges in developing scientific paper recommender systems:
Cold Start: Difficulty in recommending items to new users or recommending newly added items due to insufficient data.
Sparsity: The user-item interaction matrix is often very sparse, making it hard to find meaningful patterns or neighbours, particularly challenging for CF.
Scalability: Handling the ever-increasing number of users and papers in digital libraries and adapting to dynamic datasets.
Privacy: Protecting users' sensitive information collected for personalization while still providing effective recommendations.
Serendipity: Balancing the recommendation of highly relevant papers with introducing users to novel or unexpected relevant research that they might not have found otherwise.
Unified Scholarly Data Standards: The lack of consistent data formats and standards across different academic platforms and datasets makes data integration and model building challenging.
In conclusion, the paper provides a structured overview of the field of scientific paper recommendation, detailing the core methods, how they are evaluated, and the key challenges that researchers and practitioners face in building effective systems.