Hybrid Semantic-Graph Clustering: Enhancing Customer Segmentation with LLMs and Graph Neural Networks

Customer segmentation is a critical component of datadriven marketing and strategic decision-making. Conventional clustering algorithms such as K-Means, while computationally efficient, are inherently limited in capturing semantic nuances and relational dependencies within consumer data. This study proposes a hybrid segmentation framework that integrates Large Language Models (LLMs), specifically Sentence-BERT, with Graph Neural Networks (GNNs) using a Neo4j-based graph storage backend. The methodology follows a three-stage pipeline: (i) baseline clustering using K-Means and Principal Component Analysis (PCA), (ii) semantic clustering via Sentence-BERT embeddings combined with UMAP and HDBSCAN, and (iii) graph-based segmentation using Graph Convolutional Networks (GCNs) trained on a customer similarity graph constructed from cosine similarity metrics. Experimental evaluation on the CC GENERAL dataset comprising 9,000 credit card users demonstrates that the proposed approach significantly outperforms traditional methods. The GCN-based model achieves a Silhouette Score of 0.81, surpassing both SBERT-based clustering (0.68) and K-Means (0.52). Furthermore, it records a lower Davies-Bouldin Index (0.89) and a higher Modularity score (0.73), indicating enhanced intra-cluster cohesion and community structure. The results substantiate the efficacy of semantic-graph integration in deriving interpretable, cohesive, and business relevant customer segments.