Social networking is a technology that has been booming in recent years. With its vigorous development, various methods of predicting social networks have emerged. Most of these methods use static social networks to make predictions. However, social networks in reality are mostly dynamically changing. This research uses the data of the Mobile01 forum as the database, uses the EPMiner data mining algorithm to find frequent sequences from the data, and uses the long short-term memory (LSTM) model to predict social network relationships. This research proposes a two-tier architecture model with three-LSTM to make predictions.