Network position and scientific recognition: author centrality and citation outcomes in top AI conferences

Abstract This study investigates how authors' structural positions within scientific collaboration networks drive citation disparities beyond research content. We analyze 17,942 papers from three top AI conferences (NeurIPS, ICLR and ICML) spanning 2005 to 2024. To address the limitations of existing metrics, we introduce a novel centrality measure, HCTCD, which incorporates temporal decay and collaboration intensity. To isolate the effect of network structure, we rigorously control for textual content via a pre-trained transformer encoder. The relationship is then modeled using Beta regression, with the citation percentile within each publication year as the dependent variable. Our results demonstrate that long-term centrality exerts a stronger influence than short-term metrics. Critically, team-level centrality aggregation outperforms author-rank approaches. Furthermore, after controlling for content and other variables, incorporating a co-author with centrality 50\% higher than the first author is associated with a significant increase in expected citations, illustrating a clear “social lift”. These findings underscore the primacy of collective network connectivity and advocate for network-aware evaluation to mitigate structural inequities in scientific recognition.