Multi-View Clustering Algorithms for Big Data Analytics

Abstract

Multi-view clustering has emerged as a powerful paradigm for uncovering latent group structures in Big Data by simultaneously leveraging multiple complementary representations (views) of the same underlying entities. Traditional single-view clustering methods often suffer when confronted with heterogeneous, high-dimensional data typical of modern applications such as social network analysis, bioinformatics, and multimedia retrieval. In this manuscript, we compare and analyze three representative multi-view clustering algorithms—multi-view k-means, co-regularized spectral clustering, and deep multi-view clustering via autoencoders—on synthetic and real‐world large-scale datasets. We introduce a systematic evaluation framework that assesses clustering quality using standard validity indices (Silhouette Score, Dunn Index, Davies–Bouldin Index) and computational efficiency in terms of runtime and memory consumption.

A synthetic dataset of 10,000 samples with three distinct feature views is generated to facilitate controlled experiments, while a real-world dataset from social media image annotations is used to validate practical applicability. Our results indicate that deep multi-view clustering provides superior cluster cohesion and separation at the expense of higher computational cost, whereas co-regularized spectral clustering strikes a balance between performance and scalability. We conclude with recommendations for algorithm selection in various Big Data contexts and outline directions for enhancing scalability and robustness in future research.