<p>Brain network classification is almost entirely dependent on the original brain network topology. Due to the specificity of EEG and the uncertainty of brain network structure, the topology of primitive brain network often contains some noise edges, which reduces the generalization of classification model to a certain extent. How to realize the self-evolution of the brain network structure in the process of brain network classification is a valuable study. This study proposes a graph structure self-evolution model based on multi-task learning (GSML), which includes gated graph information fusion modules (GGIF) and graph structure self-evolution modules (GSSE). The GGIF can efficiently fuse global features and local features of brain networks, enhancing the graph perception capabilities of GSML. The function of GSSE is to predict the edges of brain networks in the process of brain network classification. In order to predict whether there are edges between nodes in a data-driven manner, GSSE uses the prior brain network structure to learn the distribution of edges in a brain network. Through the multi-task learning mechanism, the brain network classification and edge prediction tasks are integrated to enhance the brain network feature perception ability and generalization performance of GSML. Validation experiments show that: (1) Compared with the original brain network topology, GSML with self-evolution brain network structure has stronger generalization and stability; (2) GGIF can effectively enhance the graph perception ability of GSML; (3) The self-evolved brain network obtained by GSSE has better separability than the original brain network, which improves the generalization performance of GSML.</p>