Abstract
In the era of rapidly expanding social networks, community detection within social graphs plays a pivotal role in various applications such as targeted marketing, content recommendations, and understanding social dynamics. Community detection problem consists of finding a strategy for detecting cohesive groups, based on shared interests, choices, and preferences, given a social network where nodes represent users and edges represent interactions between them. In this work, we propose a hybrid method for the community detection problem that encompasses both traditional tree search algorithms and deep learning techniques. We begin by introducing a beam-search algorithm with a modularity-based agglomeration function as a foundation. To enhance its performance, we further hybridize this approach by incorporating DeepWalk embeddings into the process and leveraging a novel similarity metric for community structure assessment. Experimentation on both synthetic and real-world networks demonstrates the effectiveness of our method, particularly excelling in small to medium-sized networks, outperforming widely adopted methods.