Heterogeneous Graph Network Embedding for Sentiment Analysis on Social Media

Nowadays, more people are used to express their attitudes on different entities in online social networks, forming user-to-entity sentiment links. These sentiment links imply positive or negative semantics. Most of current user sentiment analysis literature focuses on making a positive, neutral, or negative sentiment decision according to users' text descriptions. Such approach, however, often fails to retrieve users' hidden real attitudes. We design a powerful sentiment link analysis framework named graph network embedding for sentiment analysis (NESA). NESA first utilizes variational auto-encoder (VAE) to learn joint representations of users' social relationship by preserving both the structural proximity and attribute proximity. Then, a multi-view correlation learning-based VAE is proposed to fuse the joint representation and the user-entity sentiment polarity network. By jointly optimizing the two components in a holistic learning framework, the embedding of network node information and multi-network contents is integrated and mutually reinforced. The first experimental results verify the effectiveness of adopting user, entity attributes, and social relationships for sentiment link analysis. Then we demonstrate the superiority of NESA over state-of-the-art network embedding baselines on link prediction. The last experimental results further validate that NESA model outperforms the traditional text-based sentiment prediction methods. We propose to perform sentiment analysis from network perspective; the proposed NESA model applies heterogeneous graph network embedding to fuse multi-networks information with considering their correlations and then to retrieve users' hidden real attitudes in social networks. It provides a novel angle to resolve the sentiment analysis problem.