Existing studies have proved that the resting state brain function hypernetwork based on connecting multiple nodes has been successfully applied to the diagnosis and classification of brain diseases. However, most of the hypernetworks are modeled by the Lasso method, but this method cannot solve the problem of brain function collaboration. Hence, the construction method of the hypernetwork based on group is proposed, but there are still limitations: on the one hand, they are mostly based on automatic grouping, ignoring the complex prior group structure in human brain; On the other hand, although they realize the consideration of preset groups at group level, they still choose automatically among groups, which leads to the lack of prior knowledge interpretation. Considering this problem, this paper proposes a 4-step construction of an improved the hypernetwork with a tree structured group lasso method: First, constructing the hypernetwork with the tree structured group lasso method; Secondly, calculating the local clustering attributes of the constructed hypernetwork. Then the most discriminative features are selected from the features extracted from the hypernetwork. At last, classifying the selected differences. The experimental results have shown that the tree structured group lasso method has good classification performance with an accuracy of 87.25%.