The proliferation of social networks has led users to engage across multiple platforms, creating a critical need to link their identities for comprehensive user profiling and personalized recommendations. Existing approaches face semantic discrepancies caused by platform-specific functions, interference from structurally similar but non-aligned nodes, many-to-one matching in similarity-based methods, and high computational complexity in Optimal Transport (OT)-based approaches. To overcome these issues, we propose a novel Disentanglement-Based Contrastive Learning and Optimization for user identity linkage. DCLO first disentangles user representations into network-specific and network-shared features while preserving semantic integrity with reconstruction constraints. A cross-network fusion layer integrates shared features of aligned users while aggregating neighborhood and network-specific features. DCLO employs cross-network and intra-network contrastive learning to enhance alignment and discrimination. Furthermore we design a unified node consistency transport cost that jointly captures semantic and structural consistency to optimize alignment inference. Extensive experiments on three real-world social network datasets demonstrate that DCLO consistently outperforms ten state-of-the-art baselines, achieving superior effectiveness, robustness, and generalizability.