Multi-View Clustering Model Based on Self-Supervised Comparative Learning for Bank Customer Segmentation and Risk Assessment

In the context of digital transformation in the banking industry, the high-dimensional sparsity, inter-class confusion and feature clustering target mismatch of heterogeneous customer data from multiple sources pose serious challenges to customer segmentation. Traditional methods rely on manual feature engineering or a cut-and-dried two-stage optimization process, which makes it challenging to capture the nonlinear dynamic features of customer behavior. To this end, this paper proposes a Self-supervised Contrastive Multi-View Clustering (SCMVC) model based on self-supervised comparative learning, which achieves a synergistic evolution of feature representations and clustering objectives through end-to-end joint optimization. The model takes the multi-view self-encoder as the core, constructs nonlinear dimensionality reduction and interactive feature enhancement modules to solve the problem of heterogeneous data fusion, designs the learning clustering center and t-distribution soft allocation strategy to optimize the potential spatial structure dynamically, innovatively introduces the self-adaptive view weighting mechanism, dynamically adjusts the contribution degree of multi-source data according to the real-time clustering quality, and suppresses the interference of low-quality view; combines with the comparative learning constraints to achieve the synergy of the clustering goal and feature representation through the inter-class optimization. Combined with contrast learning constraints, the discriminative properties are enhanced by cross-view embedding alignment and inter-class feature separation. Experiments based on a commercial bank's customer transaction and service data show that SCMVC significantly outperforms benchmark models such as DEC, IDEC and DEMVC in clustering accuracy, normalized mutual information and adjusted Rand index. The ablation experiments validate the central role of comparative learning and adaptive weighting. The method provides a new paradigm for intelligent customer management, and its modular design can be extended to scenarios such as risk prevention and control, precision marketing, etc. It provides theoretical methods and practical tools to support the transformation of data value in the financial industry.