Embedding-Based Clustering Framework for Categorical Survey Analysis

Clustering categorical survey data is challenging due to the absence of intrinsic ordering and complex attribute inter-dependencies. Traditional clustering algorithms often fail to achieve satisfactory performance on high-dimensional, sparse categorical datasets. To address these limitations, this work proposes a novel framework that leverages large language models (LLMs) to transform survey responses into high-dimensional semantic embeddings, enabling the application of conventional numerical clustering techniques. Experiments conducted on a real-world water consumer survey dataset demonstrate that the embedding-based approach significantly outperforms traditional one-hot encoded clustering methods. Among the tested techniques, K-means achieved the best overall performance, validated through Silhouette Coefficient, Davies-Bouldin Index, and Calinski-Harabasz Index metrics. The findings highlight the effectiveness of semantic representation learning for enhancing clustering quality in categorical survey data analysis.