Beyond Single-Score Ranking: Facet-Aware Reranking for Controllable Diversity in Paper Recommendation

Duan Ming Tao

cs.IRMar 11, 2026

Beyond Single-Score Ranking: Facet-Aware Reranking for Controllable Diversity in Paper Recommendation

Authors: Duan Ming Tao

Organizations: EECS International Graduate Program, National Yang Ming Chiao Tung University

Abstract

Current paper recommendation systems output a single similarity score that mixes different notions of relatedness, so users cannot specify why papers should be similar. We present SciFACE (Scientific Faceted Cross-Encoder), a reranking framework that models two independent facets: Background (what problem is studied) and Method (how it is solved). SciFACE trains two separate cross-encoders on 5,891 real seed-candidate paper pairs labeled by GPT-4o-mini with facet-specific criteria and validated against human judgments. On CSFCube, SciFACE reaches 70.63 NDCG@20 on Background (5.9 points above SPECTER) and 49.06 NDCG@20 on Method (31.1 points above SPECTER), competitive with state-of-the-art results. Compared with FaBLE without citation pre-training, SciFACE improves Method NDCG@20 by 4.1 points while using 5,891 labeled pairs versus 40K synthetic augmentations. These results show that high-quality grounded facet labels can be more data-efficient than large-scale synthetic augmentation for learning fine-grained scientific similarity.

Figures & tables

Figure 3.1: System overview of our facet-aware reranking pipeline. Stage 1 constructs facet-labeled triplets from seed-candidate pairs using GPT-4o-mini (Background and Method scores). Stage 2 trains two independent SciBERT cross-encoders (one for Background, one for Method) using pairwise margin ranking.

Domain	Seeds
Graph Neural Networks	56
Computer Vision	44
NLP / Text	42
Reinforcement Learning	22
Other (Robotics, Medical AI, etc.)	38
Total	202

Table 1: Distribution of seed papers by domain.

Score	Background (WHAT problem?)	Method (HOW solved?)
3	Same exact task/objective	Same core architecture (domain-independent)
2	Related tasks in same sub-area	Same method family, different architecture
1	Same ML problem type, diff domain	Specific shared mechanism
0	Unrelated communities	Different approaches OR not stated

Table 2: Facet Labeling Criteria

Comparison	Background	Method
Human vs. GPT-4o-mini (Ours)	0.61	0.44
CSFCube Pre-Adjudication	0.45	0.31

Table 3: Annotation Agreement (Spearman

\rho

)

Score	BG Count	BG %	MT Count	MT %
0	408	6.9	1,075	18.2
1	2,136	36.3	1,837	31.2
2	1,954	33.2	2,705	45.9
3	1,393	23.6	274	4.7

Table 4: Score Distribution Across All Labeled Pairs (n=5,891)

Model	Background		Method
Model	NDCG%20	MAP	NDCG%20	MAP
SPECTER Cohan et al. (2020)	66.70	43.95	37.41	22.44
+FaBle Do et al. (2025)	67.38	42.66	44.97	25.98
TSASPIRE † Mysore et al. (2022)	70.22	49.58	48.20	28.86
OTASPIRE † Mysore et al. (2022)	71.04	50.56	46.46	27.64
SPECTER-COCITE ${}_{\text{Spec}}$	70.03	49.99	45.99	25.60

Table 5: SciFACE vs. Baselines on CSFCube (Background & Method facets). All baseline results are reproduced from FaBLE.

\dagger

denotes models initialized from SPECTER-COCITE

{}_{\text{Spec}}

, a bi-encoder pre-trained on 1.3M co-citation pairs. SciFACE uses only 5,891 GPT-labeled real pairs with no citation-based pre-fine-tuning.

Query / Candidate Pair	GT	Our Rank
Query: Edge-Linear First-Order Dependency Parsing with Undirected MST Inference Candidate: Finding Optimal Bayesian Network Given a Super-Structure	2	#87/250

Table 6: Granularity Mismatch Example

Query / Candidate Pair	GT	Our Rank
Query: Compact Personalized Models for Neural Machine Translation Candidate: Direct adaptation of hybrid DNN/HMM model for fast speaker adaptation	2	#92/96

Table 7: Cross-Domain Terminology Barrier Example

Scale	NDCG%20
0–3 (original)	45.57
0–2 (merged)	49.06

Table 8: Method Facet: Label Scale Ablation (CSFCube Test Set)

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Background	Method
Human vs. GPT-4o-mini (Ours)	0.61	0.44
CSFCube Pre-Adjudication	0.45	0.31

Appendix

Table 9: Spearman

\rho

Agreement Comparison

Score	Background		Method
Score	Human	GPT	Human	GPT
0	32	5	67	20
1	31	51	20	19
2	24	23	5	58
3	13	21	8	3

Appendix

Table 10: Score Distribution Comparison (n=100)

Human $\backslash$ GPT	0	1	2	3
0	5	24	3	0
1	0	18	9	4
2	0	5	11	8
3	0	4	0	9

Appendix

Table 11: Background Confusion Matrix (Human

\times

GPT-4o-mini)

Human $\backslash$ GPT	0	1	2	3
0	20	15	32	0
1	0	2	18	0
2	0	0	5	0
3	0	2	3	3

Appendix

Table 12: Method Confusion Matrix (Human

\times

GPT-4o-mini)

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 502 papers

GRAFT: Graph-Distilled Generative Retrieval for Facet-Aware Scientific Literature Exploration
Aug 23, 2026Italo Luis da Silva, Hanqi Yan, Yujing Wang +3Information Retrieval Generative Retrieval
Department of Informatics King’s College London London, England, UK
From Isolated Scoring to Collaborative Ranking: A Comparison-Native Framework for LLM-Based Paper Evaluation
Mar 18, 2026Pujun Zheng, Jiacheng Yao, Jinquan Zheng +6Automated Peer Review Scholarly Peer Review
School of Economics and Management, East China Normal University · School of Information Management, Wuhan University · China Academic Degrees & Graduate Education Development Center
PaperFlow: Profiling, Recommending, and Adapting Across Daily Paper Streams
Jun 5, 2026Fuqiang Wang, Song Tan, Zheng Guo +8Information Retrieval Recommender Systems
1Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center (National Supercomputer Center in Jinan), Qilu University of Technology (Shandong Academy of Sciences) · University of Chinese Academy of Science · 3Shanghai Artificial Intelligence Laboratory

Aug 23, 2026J/K move · Enter open · S save

GRAFT: Graph-Distilled Generative Retrieval for Facet-Aware Scientific Literature Exploration

Italo Luis da Silva, Hanqi Yan, Yujing Wang, Jiangnan Ye+2

Department of Informatics King’s College London London, England, UK

Information Retrieval Generative Retrieval Graph-Based Retrieval

Open paper AskPDF arXiv

Scientific papers may relate by problem, method, result, or contribution, but document-level retrievers collapse these into a single similarity score without saying why they are related. Citation- and similarity-based retrieval alone also confines search to the neighbourhood of what is already known, whereas generative retrieval generates document identifiers directly, enabling the exploratory retrieval that scientific discovery depends on. We connect papers in a graph whose edges are typed by these four facets, derived from facet items and citation signals, and distil it into a generative retriever whose identifiers are the papers' own facet text. Two graph properties do not survive naive distillation. First, because every training pair is an edge, naive enumeration indexes just 84% of the corpus. Coverage-aware distillation makes every paper learnable through a reverse-neighbour fallback, a minimum-coverage threshold, and edge-importance weighting. Second, constrained decoding guarantees that every generated identifier is a valid paper, but not that the graph connects it to the query. Graph-weighted reciprocal rank fusion scales each candidate's rank term by its query-candidate edge weight, dropping unsupported ones. On LitWeave, our constructed corpus of 11,359 NLP papers, Graft recovers 91% of its graph teacher's Recall@20 with no nearest-neighbour index or encoder at inference, and outperforms the graph teacher on query papers outside the corpus. It reproduces the graph's own facet labels at 0.922 precision, so every returned paper arrives labelled with the facet that surfaced it rather than an opaque score.

Abstract

Recommender Systems

Explore similar work

GRAFT: Graph-Distilled Generative Retrieval for Facet-Aware Scientific Literature Exploration

From Isolated Scoring to Collaborative Ranking: A Comparison-Native Framework for LLM-Based Paper Evaluation

PaperFlow: Profiling, Recommending, and Adapting Across Daily Paper Streams