Mitigating data imbalance in scholarly networks using topic models and social network analysis

Representing researchers in large-scale scholarly environments is increasingly challenging due to the rapid growth and interdisciplinary nature of scientific publication. Most existing approaches rely on observable relationships such as co-authorship or citation links, which often reinforce established connections and overlook latent topical structure. In this work, we construct topic-based scholarly networks using publication content alone, representing researchers through topic distributions derived from both probabilistic and embedding-based topic models, including LDA and BERTopic, to compare different modeling paradigms. A central challenge is publication imbalance, where dominant research themes of prolific authors overshadow their secondary or emerging interests. When publication content is aggregated at the researcher level, these dominant themes disproportionately shape the resulting representation, distorting similarity estimates and downstream community structure. To address this issue, we introduce a cloning-based representation that partitions a researcher’s publications into topic-specific components and represents each component as an independent node in the network. This enables researchers to participate in multiple topical communities and naturally induces overlapping community structure. Empirical analysis on institutional publication data, together with validation against co-authorship patterns, suggests that accounting for publication imbalance improves the alignment of topic-based representations with observed collaborative relationships and reveals latent interdisciplinary structure that is often obscured under standard aggregation.