Academic Search is a search task aimed to manage and retrieve scientific documents like journal articles and conference papers. Personalization in this context meets individual researchers' needs by leveraging, through user profiles, the user related information (e.g. documents authored by a researcher), to improve search efectiveness and to reduce the information overload. While citation graphs are a valuable means to support the outcome of recommender systems, their use in personalized academic search (with, e.g. nodes as papers and edges as citations) is still under-explored.
https://pkasela.github.io/ (P. Kasela); http://raffaele.isti.cnr.it/ (R. Perego) CEUR
ceur-ws.org
QUERY
Top-k
Neural LM Top-k
Query Embedding
Doc Embeddings Top Authors Embeddings DenseScSoimreilarity
First Stage
Score User Similarity
Score KG Embedding
User Embedding
Final Score
The neural bi-encoder model is trained by minimizing the distance between the query representation and the associated relevant document representations while increasing the distance between the query representation and the non relevant documents representations using the Triplet Margin Loss [16]. The final score for each document is given by a weighted combination of: • BM25(, ) : lexical similarity. • Dense(, ) : semantic similarity from the neural model. • UserSim(, ): cosine similarity between the user embedding and the authors of document in the knowledge graph embedding.
(, ) = 1 ⋅ BM25(, ) + 2 ⋅ Dense(, ) + 3 ⋅ UserSim(, ) (1) with 1 + 2 + 3 = 1 and are optimized on a validation split.
PARK encodes the authors in a vector space shared with the language model in order to capture both bibliographic structure and textual semantics.
Academic Knowledge Graph Construction Starting from a citation graph (papers as nodes; citations as directed edges), we build a heterogeneous Knowledge Graph (KG) with four node types and ifve relations: • Nodes: Author (user), Document (paper), Venue (conference or journal), and Afiliation (institution) • Relations: wrote: Author → Document, cited: Document → Document, co_author: Author ↔ Author, in_venue: Document → Venue, and afiliated: Author → Afiliation
DocAuumthoernt1
Ve.n.ue Cited
Cited DocAuumthoernt2 Cited DocAuumthoernt3
Ve.n.ue Ve.n.ue
Venue2 Affiliation2
In_Venue In_Venue
Affliliated Venue1
Co_author Wrote Knowledge Graph
Affiliation1 In_Venue Affiliated
USER1
Cited Document1 USER2 Wrote
Document2 Venue1 Venue2
User1 Afiliation1 User2
Afiliation2
We embed the KG into the same -dimensional latent space as our neural
1. Document nodes are initialized with fixed embeddings from the pre-trained MiniLM encoder.
2. Other nodes (authors, venues, afiliations) and all relations are jointly embedded using: TransE
[
User Embeddings & Scoring Each author is represented by their learned KG embedding. At query time, we compute the user similarity score as the cosine similarity between the user embeddings of the query issuer and the authors of the documents being scored. The score reflects the alignment of the research profiles and interests of the user issuing the query and the authors of the paper being scored. This score is integrated into our final ranking formula (Eq. 1).
The system is evaluated on four datasets specifically designed for evaluating model in the context
of personalized academic search (Computer Science, Political Science, Psychology, and Physics) [17].
We compare against the following baselines: BM25 [14], MiniLM [15], Mean [17], Attention [18],
Self Citation, CrossEncRA [
PARK outperforms all baselines in Political Science, Psychology, and Physics (Table 1). In Computer Science, the POP baseline remains competitive due to the high predictive value of citation counts. Overall, PARK demonstrates robust efectiveness across diverse disciplines.
To understand the contribution of node types in the knowledge graph, we conduct an ablation study using PARK-H (Table 2). The goal was to evaluate the efect of each node type (user, venue, and afiliation) and their associated relations on the model’s retrieval performance. Results in Table 2 indicate:
• Using only user nodes yields substantial gains over baseline. 1https://github.com/pkasela/PARK-Personalized_Academic_Retrieval_with_Knowledge-graphs Efectivess of PARK-E and PARK-H compared to the competing methods on the four datasets. The best-performing model is highlighted in boldface. Symbol * indicates a statistically significant diference over the second-bestperforming model. Computer Science
Political Science
Psychology 0.123 0.193 0.199 0.201 0.213 0.213 0.238* 0.228 0.230
BM25 MiniLM
Mean
Attention Self Citation
CTRLIt CrossEncRA PageRank
POP PARK-E
PARK-H
PARK demonstrates how knowledge graph-based user embeddings, when aligned with neural document encoders, can improve personalized academic search. By representing academic entities in a unified latent space, PARK captures both explicit citation relationships and latent author-topic patterns.
While PARK performs well across most domains, limitations remain. The use of fixed document embeddings may limit adaptability, and citation coverage biases could afect robustness. Future work will explore: softening constraints on document embeddings; integrating popularity-based priors for domains like CS; and adapting to open-world or streaming academic corpora.
Overall, PARK advances the state-of-the-art in personalized academic search by combining structured knowledge with dense text representations.
This work was supported by the European Union – Next Generation EU within the project NRPP M4C2, Investment 1.,3 DD. 341–15 march 2022 – FAIR – Future Artificial Intelligence Research – Spoke 4 PE00000013 - D53C22002380006.
This work was partially supported by the Spoke “Human-centered AI” of the M4C2 - Investimento 1.3, Partenariato Esteso PE00000013 - “FAIR - Future Artificial Intelligence Research”, the “Extreme Food Risk Analytics” (EFRA) project, Grant no. 101093026, funded by European Union – NextGenerationEU.
During the preparation of this work, the author(s) used GPT-4 in order to: Grammar and spelling check, Paraphrase and reword. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.
Ablation study results for each node type on four datasets. Proceedings of the AAAI Conference on Artificial Intelligence 28 (2014). doi: 10.1609/aaai.v28i1. 8870. [14] S. E. Robertson, S. Walker, S. Jones, M. Hancock-Beaulieu, M. Gatford, Okapi at TREC-3, in: D. K.
Harman (Ed.), Proceedings of The Third Text REtrieval Conference, TREC 1994, Gaithersburg, Maryland, USA, November 2-4, 1994, volume 500-225 of NIST Special Publication, National Institute of Standards and Technology (NIST), 1994, pp. 109–126. URL: http://trec.nist.gov/pubs/trec3/ papers/city.ps.gz. [15] W. Wang, F. Wei, L. Dong, H. Bao, N. Yang, M. Zhou, Minilm: Deep self-attention distillation for task-agnostic compression of pre-trained transformers, in: H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, H. Lin (Eds.), Advances in Neural Information Processing Systems, volume 33, Curran Associates, Inc., 2020, pp. 5776–5788. URL: https://proceedings.neurips.cc/paper_files/paper/2020/ file/3f5ee243547dee91fbd053c1c4a845aa-Paper.pdf. [16] V. Balntas, E. Riba, D. Ponsa, K. Mikolajczyk, Learning local feature descriptors with triplets and shallow convolutional neural networks., in: Bmvc, volume 1, 2016, p. 3. [17] E. Bassani, P. Kasela, A. Raganato, G. Pasi, A multi-domain benchmark for personalized search evaluation, in: Proceedings of the 31st ACM International Conference on Information & Knowledge Management, CIKM ’22, Association for Computing Machinery, New York, NY, USA, 2022, p. 3822–3827. doi:10.1145/3511808.3557536. [18] J. Jiang, T. Wu, G. Roumpos, H. Cheng, X. Yi, E. Chi, H. Ganapathy, N. Jindal, P. Cao, W. Wang, End-to-end deep attentive personalized item retrieval for online content-sharing platforms, in: WWW ’20: The Web Conference 2020, Taipei, Taiwan, April 20-24, 2020, ACM / IW3C2, 2020, pp. 2870–2877. [19] S. Mysore, G. Dhanania, K. Patil, S. Kallumadi, A. McCallum, H. Zamani, Memory augmented cross-encoders for controllable personalized search, 2024. URL: https://arxiv.org/abs/2411.02790. arXiv:2411.02790. [20] L. Page, S. Brin, R. Motwani, T. Winograd, The PageRank Citation Ranking: Bringing Order to the Web., Technical Report 1999-66, Stanford InfoLab, 1999. URL: http://ilpubs.stanford.edu:8090/422/, previous number = SIDL-WP-1999-0120.