Public Profile Matters: A Scalable Integrated Approach to Recommend Citations in the Wild
New system captures human citation patterns without bias, achieving state-of-the-art results with a novel inductive evaluation.
A research team has published a paper introducing a significant advancement in AI-powered citation recommendation systems. The work, titled 'Public Profile Matters: A Scalable Integrated Approach to Recommend Citations in the Wild,' addresses two major flaws in current methods: their neglect of nuanced human citation behavior and their unrealistic evaluation protocols. The researchers propose a lightweight, non-learnable module called 'Profiler' that efficiently captures patterns in how authors cite each other based on their public academic profiles, doing so without introducing the computational cost or systematic bias seen in recent methods.
Furthermore, the team identified that existing systems are typically tested in a 'transductive' setting, where a model is evaluated on papers it was trained on, which doesn't reflect the real-world task of suggesting citations for brand new research. They introduced a rigorous 'inductive' evaluation with strict temporal constraints to simulate this authentic scenario. Finally, they presented 'DAVINCI,' a novel reranking model that uses an adaptive vector-gating mechanism to combine the confidence scores from the Profiler with deep semantic information from the paper text.
The integrated system demonstrates new state-of-the-art performance across multiple benchmark datasets, showing superior efficiency and generalizability. By making the retrieval of candidate papers more accurate and the evaluation more realistic, this research provides a more robust framework for building tools that can genuinely assist researchers in the literature review and paper-writing process.
- Introduced 'Profiler,' a lightweight module that captures human citation patterns from public academic profiles without learnable parameters or bias.
- Proposed a new 'Inductive' evaluation setting with temporal constraints, moving beyond flawed 'transductive' tests to better simulate real-world use.
- Developed the 'DAVINCI' reranker, which integrates Profiler's signals with semantic data via adaptive gating, achieving SOTA results on benchmarks.
Why It Matters
Provides a more accurate, efficient, and realistic framework for AI tools that help researchers find and cite relevant literature.