Drastic Fluctuation Prediction of the Number of Comments on Social Media by Attributes of Comments

Abstract

Online social media has an exponential level of communication speed in terms of message dissemination. Users can publish comments freely to various web content on a characteristic network of communicators and viewers. Many of these comments contain emotions or opinions of users, which may cause sympathy and influence others’ comments. Moreover, such comments may raise social responses, i.e. they may cause drastic fluctuations in the number of comments. In this study, using the content of textual comments, we propose two structural approaches (PDFCPL and PDFCML) to predict the future drastic fluctuation in the number of comments based on Long Short-Term Memory (LSTM). To quantify each textual comment, we define two attributes: (1) relevance to its relevant topic based on cosine similarity and (2) importance of its content which is calculated by TF-IDF. The predictions are made by these attributes and the number of previously observed comments as well. To evaluate the performance of our approaches, we conduct comparing experiments with other methods on real data of Twitter. The results present that the proposed method PDFCPL has better performance than existing methods to predict the occurrence of drastic fluctuation in the number of comments.