Academic Article Recommendation Using Multiple Perspectives

TL;DR

This paper argues that Content-Based Filtering (CBF) and Graph-Based (GB) methods deliver complementary signals for academic paper recommendations by treating the literature as a dialogue between authors and readers. It uses Specter (CBF) and ProNE (GB) as representative embeddings and analyzes nine dimensions where their behavior and limitations diverge, outlining practical hybrid opportunities. The authors discuss inputs, interpretations, history, computation, scale, time dynamics, priors, and corner cases to show how combining the two perspectives can improve coverage and robustness, especially on large Semantic Scholar-scale graphs. The work has practical implications for delivering more comprehensive related-work suggestions, organizer tools, and explainable hybrid systems in scholarly search.

Abstract

We argue that Content-based filtering (CBF) and Graph-based methods (GB) complement one another in Academic Search recommendations. The scientific literature can be viewed as a conversation between authors and the audience. CBF uses abstracts to infer authors' positions, and GB uses citations to infer responses from the audience. In this paper, we describe nine differences between CBF and GB, as well as synergistic opportunities for hybrid combinations. Two embeddings will be used to illustrate these opportunities: (1) Specter, a CBF method based on BERT-like deepnet encodings of abstracts, and (2) ProNE, a GB method based on spectral clustering of more than 200M papers and 2B citations from Semantic Scholar.

Figures (5)

• Figure 1: $|A \cap L| \approx 30\%$
• Figure 2: ProNE prefactorization (SVD) takes a few days and a few TBs on larger graphs.
• Figure 3: Larger graphs favor ProNE (GB).
• Figure 4: GB recommendations have more citations, but are less recent.
• Figure 5: The spike near 1 (top) is dominated by corner cases involving abstracts. ProNE (bottom) is not bimodal.