Public Profile Matters: A Scalable Integrated Approach to Recommend Citations in the Wild
Summary: arXiv:2603.17361v2 Announce Type: replace-cross
Abstract
Proper citation of relevant literature is essential for contextualising and validating scientific contributions. While current citation recommendation systems leverage local and global textual information, they often overlook the nuances of human citation behaviour. Recent methods that incorporate such patterns improve performance but incur high computational costs and introduce systematic biases into downstream rerankers. To address this, we propose Profiler, a lightweight, non-learnable module that captures human citation patterns efficiently and without bias, significantly enhancing candidate retrieval.
Introduction
The importance of accurate citation in scientific literature cannot be overstated. Citations not only lend credibility to research but also provide a pathway for future exploration and development. Yet, many citation recommendation systems struggle to effectively mirror the complexities of human citation behaviour, leading to inefficiencies and biases.
Methodology
To overcome the limitations of existing systems, our research introduces Profiler, which serves as a non-learnable module designed to efficiently capture human citation patterns. This approach aims to enhance the candidate retrieval process by addressing the following key areas:
- Efficiency: Profiler operates without the computational overhead associated with traditional learning methods.
- Bias Reduction: By not relying on learnable parameters, Profiler minimizes the introduction of biases that often affect citation recommendation systems.
- Enhanced Candidate Retrieval: The integration of human citation patterns results in a more accurate representation of potential citations.
Evaluation Protocols
In addition to introducing Profiler, our research identifies a significant limitation in current evaluation protocols. Most systems are assessed in a transductive setting, which does not accurately reflect real-world scenarios where citations must be recommended for newly authored papers. To address this, we propose a rigorous inductive evaluation setting that enforces strict temporal constraints.
The DAVINCI Model
Alongside Profiler, we present DAVINCI, a novel reranking model that leverages profiler-derived confidence priors. DAVINCI integrates semantic information through an adaptive vector-gating mechanism, ensuring that citation recommendations are both relevant and timely. The key features of DAVINCI include:
- Integration of Human Patterns: Utilizes insights derived from Profiler for enhanced recommendation accuracy.
- Adaptive Mechanism: The vector-gating feature allows for dynamic adjustments based on the context of the citation.
- State-of-the-Art Performance: DAVINCI has demonstrated superior results across multiple benchmark datasets, showcasing its efficiency and generalisability.
Conclusion
Our findings indicate that incorporating a lightweight, non-learnable module like Profiler into citation recommendation systems can significantly enhance performance while mitigating biases. The introduction of a rigorous inductive evaluation setting and the development of the DAVINCI model further represent critical advancements in the field of citation recommendation. This integrated approach not only improves the accuracy of citations but also aligns with the complexities of real-world research dynamics.