This paper proposes a collaborative sequential recommendation framework that integrates both context information from individual user behavior sequences and collaborative information across different users' behavior sequences. The framework builds a local dependency graph for each item to encode these two types of information. The model uses a hierarchical graph aggregation mechanism to extract multi-view representations of user behavior sequences, which are then passed through a transformer-based model to capture both sequential and collaborative patterns. The proposed model is evaluated on five benchmark datasets and shows superior performance compared to existing sequential recommendation methods. The key contributions include the design of a hierarchical graph aggregation model, the application of multi-view representations to enhance user behavior modeling, and the introduction of a Dirichlet sampling method to improve model robustness. The model's time complexity is analyzed, showing that it can efficiently handle large-scale item dependency graphs. The framework is able to capture higher-order item dependencies and is more robust to user behavior fluctuations, leading to improved recommendation accuracy.This paper proposes a collaborative sequential recommendation framework that integrates both context information from individual user behavior sequences and collaborative information across different users' behavior sequences. The framework builds a local dependency graph for each item to encode these two types of information. The model uses a hierarchical graph aggregation mechanism to extract multi-view representations of user behavior sequences, which are then passed through a transformer-based model to capture both sequential and collaborative patterns. The proposed model is evaluated on five benchmark datasets and shows superior performance compared to existing sequential recommendation methods. The key contributions include the design of a hierarchical graph aggregation model, the application of multi-view representations to enhance user behavior modeling, and the introduction of a Dirichlet sampling method to improve model robustness. The model's time complexity is analyzed, showing that it can efficiently handle large-scale item dependency graphs. The framework is able to capture higher-order item dependencies and is more robust to user behavior fluctuations, leading to improved recommendation accuracy.