The consumption of high doses of marijuana can have significant psychological and social impacts. In this study, we propose an interpretable novel framework called the HOGAB (High-Order Graph Attention Neural Networks) model for addictive Marijuana classification and analysis of the localized network clusters that demonstrated abnormal brain activities among chronic marijuana users. The HOGAB integrates dynamic intrinsic functional networks with LSTM technology to capture temporal patterns in fMRI time series of marijuana users. We employed the high-order attention module in neighborhood nodes for information fusion and message passing, enhancing community clustering analysis for long-term marijuana users. Furthermore, we improve the overall classification ability of the model by incorporating attention mechanisms, achieving an AUC of 85.1% and an accuracy of 80.7% in classification, higher than the comparison algoirthms. Specifically, we identified the most relevant subnetworks and cognitive regions that are influenced by persistent marijuana usage, revealing that chronic marijuana consumption adversely affects cognitive control, particularly within the Dorsal Attention and Frontoparietal networks, which are essential for attentional, cognitive and higher cognitive functions. The results show that our proposed model is capable of accurately predicting craving bahavior and identifying brain maps associated with long-term cravings, and thus pinpointing brain regions that are important for analysis.