Amidst the exponential increase of data in the internet era, sequential recommendation systems have been instrumental in capturing the evolving preferences of users by analyzing their historical behaviors, thereby swiftly providing users with information of interest. These systems find extensive applications in domains such as e-commerce, social platforms, music, and movie websites. However, the majority of current approaches rely on static assumptions, overlooking the dynamic of user preferences and behaviors. Furthermore, these methods typically consider only a single type of behavioral data, ignoring the fine-grained and meaningful preferences hidden within multiple behaviors (clicks, purchases, and favorites) in sequences. Addressing these limitations, we propose a Multi-Behavior Multi-Scale Time Interval (MB-MSTI) method. This paper involves a multi-scale time interval Transformer for dynamic temporal modeling, a multi-behavior-item heterogeneous graph for enriched item relationship understanding, and a graph attention mechanism for node relevance extraction. Experiments on Taobao and IJCAI datasets show our model’s superior performance over current leading methods.