SEMANTIC-AWARE POPULARITY PREDICTION OF SOCIAL MEDIA POSTS USING ROBERTA AND BILSTM NETWORKS

Abstract

The popularity of user-generated content has rapidly increased with the popularity of social media platforms, which have created many opportunities for the prediction of online post popularity. Hard-coded features and shallow neural networks, such as Random Forests, are not expressive enough to capture the semantic nature of language and the contextual phenomena of social media. To overcome these issues, we develop a semantic-aware model for popularity prediction based on a hybrid RoBERTa-BiLSTM architecture. RoBERTa extracts deep contextual semantics in post-text, and BiLSTM models temporal dependencies between tokens. The proposed system significantly outperforms baseline approaches, achieving state-of-the-art results with an accuracy of 1.00, an F1 score of 0.99, and a Spearman Rank Correlation of 0.9985. Comparative analysis with Transformer, LSTM, and BERT-based models further confirms the robustness of our model. Moreover, the MAE and MSE error metrics suggest that the proposed model effectively reduces prediction errors. Experiments were conducted on the SMPD dataset, a benchmark for social media popularity prediction. The results presented here argue for the need to use semantics and sequential modeling in practical social media applications.