In traditional bibliometrics, every citation is treated as an identical unit of currency. Whether a paper receives a passing, perfunctory mention in a literature review or serves as the foundational methodology for a groundbreaking discovery, standard metrics like raw citation counts and the $h$-index register both as exactly "+1".
As noted by Saurav Manchanda et al., this uniform weighting introduces massive noise into scientific evaluation, obscuring genuine intellectual influence. Furthermore, Upul Senanayake points out that flat counting leaves metrics highly vulnerable to manipulation, such as citation cartels or excessive self-citation. To resolve this, a major research direction focuses on non-uniform citation weighting, evaluating scholarly impact through two primary mechanisms: network-based prestige and content-based intensity.
Recursive Prestige Weighting
How do we measure the authority of a citation without manually reading every paper? The answer lies in graph topology. By treating scientific literature as a directed graph—where papers, authors, or journals are nodes and citations are directed edges—we can use recursive centrality algorithms to let prestige flow through the network.
In this framework, a citation's value depends entirely on the prestige of the citing node. For instance, Andrew Lim 0001 et al. applied eigenvector centrality to academic journals, demonstrating that recursive analysis aligns much more closely with expert human judgments of prestige than simple citation counts. Similarly, Erjia Yan et al. extended this to coauthorship networks, using a weighted PageRank to separate widespread popularity from deep-seated academic authority.
[Low-Prestige Paper] ──(+1)──► [Target Paper] ◄──(+10)── [High-Prestige Paper]
▲
│
(Uniform Weighting = 2)
(Prestige Weighting = 11)
This approach works because it models scientific impact as a mutual reinforcement process. As explored by Yujing Wang et al., papers derive prestige from the reputation of their authors and journals (acting as hubs), while the hubs' reputations are conversely updated based on the quality of the papers they produce. Crucially, Upul Senanayake demonstrates that this recursive weighting makes metrics highly resistant to gaming: to artificially inflate a score, an actor cannot simply spam low-quality self-citations; they must secure citations from sources that are themselves highly ranked by the community.
Content and Intensity Mapping
While network structure captures who is citing a paper, it cannot capture how the paper is being used. A paper might be cited once in a long list of background alternatives, or its core algorithm might be discussed across multiple sections of a text.
To capture this nuance, researchers use natural language processing (NLP) to analyze the full text of citing documents. Xiaodan Zhu 0001 et al. demonstrated that weighting citations based on internal mention counts—how many times a specific paper is referenced within the text—drastically reduces the noise of superficial citations. Taking this further, Qiaozhu Mei et al. incorporated both the authority of the citing source and the physical proximity of words to the citation anchor to improve scientific summarization, ensuring that the text most directly impacted by the reference is prioritized.
Rather than treating citations as binary connections, models like the one proposed by Tanmoy Chakraborty et al. quantify "reference intensity" on a granular scale, allowing bibliometric indices to prioritize foundational contributions over perfunctory mentions.
This distinction is even being built into modern machine learning architectures. For example, Zhentao Liang et al. introduced a contrastive learning framework for document representation where low-importance, perfunctory citations are explicitly treated as "hard negatives" during training. This forces the model to learn representations based strictly on core thematic and intellectual connections, resulting in far more precise maps of global science.