Advancing Community Detection through Ensemble Learning and Modularity Maximization

Arguably, the most fundamental problem in Network Science is finding structure within a complex network. Often, this is done by partitioning the network's nodes into communities in a way that maximizes an objective function. However, finding the maximizing partition is generally a computationally difficult NP-complete problem. Recently, a machine-learning algorithmic scheme was introduced that uses information within a set of partitions to find a new partition that better maximizes an objective function. The scheme, known as RenEEL, uses extremal ensemble learning. Starting with an ensemble of $K$ partitions, it updates the ensemble by considering replacing its worst member with the best of $L$ partitions found by analyzing a reduced network formed by collapsing nodes, which all the ensemble partitions agree should be grouped together, into super-nodes. The updating continues until consensus is achieved within the ensemble about what the best partition is. The original $K$ ensemble partitions and each of the $L$ partitions which is used for the update, are found using a simple ``base" partitioning algorithm. We conduct an empirical study of RenEEL’s effectiveness as a function of ensemble parameters and relate the results to extreme value statistics. It shows that increasing the ensemble size $K$ yields better results in approaching ensemble size $L$. Building on this foundation, we extend the RenEEL framework to the domain of bipartite networks, where community detection presents the unique challenge of the resolution limit in modularity-based methods. We first demonstrate how a benchmark bipartite network fails to resolve smaller communities using traditional modularity. We then introduce a new metric, Generalized Bipartite Modularity Density (Q_bg), which leverages the resolution limit to reveal hierarchical community structures in bipartite networks by allowing tunable control over resolution. This consistently outperforms existing metrics in detecting communities in both benchmark and real bipartite networks. Using RenEEL to maximize this novel metric, we demonstrate its effectiveness in uncovering hierarchical community structures across a range of real-world bipartite networks, including the Southern Women network, an Asthma-Patient bipartite network, and a Psychological Item–Embedding Bipartite Network. Applying this to the Southern Women network, we show that for lower values of \(\chi\), Q_bg recovers the canonical partition reported in earlier studies. As \(\chi\) increases, it uncovers progressively finer subgroups, ultimately revealing a richer hierarchy of smaller clusters that together provide, for the first time, a unified view of the network’s multi-scale social organization. In the Asthma-Patient bipartite network of patients and cytokines, Q_bg recovers known baseline groupings at standard resolution. As resolution increases, it reveals finer modules linking specific cytokines to patient subsets, suggesting potential therapeutic targets. Notably, some slightly different groupings from the standard partition also emerge, highlighting structural features that may inform asthma treatment. In the Psychological Item–Embedding Bipartite Network derived from an LLM fine-tuned on annotated survey data, Q_bg effectively captures nuanced relationships between survey items and latent model dimensions, revealing meaningful clusters that align with conceptual similarities learned by the model. These results highlight that RenEEL, when combined with problem-specific metrics such as Q_bg, offers a robust and generalizable approach for detecting multi-scale community organization in diverse bipartite systems.

Name thesis:degree_name

Doctor of Philosophy

Discipline thesis:degree_discipline

Physics

Grantor

University of Houston

Year dc:date.issued

2025

Author dc:creator

• Ghosh, Tania 1996-

Advisor dc:contributor.advisor

• Bassler, Kevin E.

Committee members dc:contributor.committeemember

• Gunaratne, Gemunu
• Josic, Kresimir
• Weglein, Arthur B.
• Morrison, Greg

• Physics

Language dc:language.iso

English

Handle dc:identifier.uri

https://hdl.handle.net/10657/20714

OAI identifier oai:identifier

oai:uh-ir.tdl.org:10657/20714