In view of the insufficient mining of implicit associative relationships and the problem of neglecting the temporal evolution factor, we propose a domain-specific social events detection method via temporal evolution feature mining. Our model divides data into slices by time, and takes duplicate event records from different sources into account which reflects the importance of events, and constructs entity interaction graph, which reduces the impact of database error. Multi-relational graph convolutional network is improved to update graph structure information of historical evolution sequence by interaction relationships. Attention mechanism is used to learn core features to obtain global embedding of sequence units. Implicit association is mined sufficiently. Based on recurrent neural network, temporal evolution features are extracted to obtain the global embedding and the temporal evolution factor is mined effectively. Experiment results show that our method can be applied to detection task, which is better than existing methods.