Enhancing leader-based community detection in scientific collaboration using publications metadata

Leader-based approaches to community detection provide a powerful framework for uncovering group structures in collaboration networks characterized by degree heterogeneity, but are typically grounded in purely structural information, which limits their ability to capture thematic cohesion in scientific communities. This work introduces an attribute-driven extension of the Density-based Community Detection (DeCoDe) framework, designed for collaboration networks where communities emerge jointly from network topology and publication metadata. Focusing on scientific collaboration among Italian scholars as a case study, we construct a network from the Italian Ministry of University and Research and Scopus data, where nodes represent authors and edges encode co-authorship ties, and we characterize each node with attributes derived from publication-level textual descriptors. We represent the scientific content of papers using their topics, which are aggregated at the author level and transformed into dense semantic embeddings via the all-MiniLM-L6-v2 SentenceTransformer model, yielding a high-dimensional attribute space tailored to quantitative text analysis. In this space, we estimate local attribute density using a density estimator to identify leaders as nodes located in high-density regions of the embedding space, rather than via structural centrality alone as in traditional leader-based approaches. These attribute-based densities are then integrated into an extended DeCoDe procedure, which defines communities as connected components in upper-level sets of the density function, thereby enforcing both structural connectivity and thematic homogeneity. Results on empirical data indicate that attribute-based density estimation supports the identification of communities that are both densely interconnected and thematically interpretable, highlighting leaders whose profiles are representative of the dominant themes of their groups.