Multi-behavior models incorporate various interaction behaviors to comprehensively understand user preferences. Recognizing that users in real-world scenarios typically follow a certain order (i.e., view → cart → purchase), recent studies have designed cascading graph convolutional networks, aiming to leverage such dependencies in the chain to assist representation learning. Despite their effectiveness, two key limitations persist: (i) Most existing cascading models ignore latent intent factors (e.g., appearance, brand, and price) behind interactions, restricting them to a singular preference enhancement process in the behavior chain, thereby failing to adapt to complex behavior patterns. (ii) Due to the message passing mechanism in GCN, inherent issues in multi-behavior data (i.e., noise in auxiliary behaviors and sparsity in target behavior) are further amplified, leading to noise accumulation in cascading networks and the homogenization of target behavior embeddings. Thus, we propose the Self-Supervised Graph-Optimized Cascading Disentangled GCNs for Multi-Behavior Recommendation (SGCD) framework. Specifically, we first design the cascading disentangled graph convolutional networks to avoid the interference of high-order connections under different intents, while simultaneously adapting to the user’s complex behavior patterns. Additionally, to improve graph modeling capability, we present a graph structure mutual optimization between auxiliary and target behaviors. Extensive experiments demonstrate that SGCD outperforms state-of-the-art baselines.