Microblog sentiment analysis via user representative relationship under multi-interaction hybrid neural networks

In microblog sentiment analysis, the similarity between microblogs and social attributes between users are often used, but previous work usually contains a large number of disruptive microblogs due to the consideration of the relationships between the target microblog user and many other users. We attempt to explore this problem from the perspective of user representative relationship and propose the multi-interaction hybrid neural networks that consist of graph attention networks, convolutional neural network, and bidirectional long short-term memory network with bidirectional gated recurrent unit and attention. Unlike existing methods for finding similar microblog sequences or adding user node embeddings, our method focuses on using the most similar user to the target microblog user rather than all similar users, which can avoid the interference of invalid microblogs of similar users. The experimental results on four real-world microblog datasets demonstrate the superiority of our method compared with nine state-of-the-art methods.