Clustered Bert Model for predicting Retweet Popularity

This work aims to predict retweet popularity of covid19 tweet corpus. Our work fuses unsupervised and supervised learning techniques to create retweet popularity model. In the first phase, we use a Clustered Bert model, which works on clustering the Bert embeddings using clustering algorithms on the textual data to generate novel and meaningful feature set for our model. In the second phase, we use the output of Clustered Bert model as an input to the Supervised Regression models intending to predict retweet popularity. Our work also draws a comparison between features from numeric model; emotions/sentiment model; and Clustered Bert model. Three different Regression models, belonging to different categories: Nearest Neighbors, Ensemble and Stacked models are then tested on the final feature-set to generate predictions for our model. The results show higher accuracy when the Clustered Bert model is used in combination with numerical and emotion/sentiment model. The experiment shows better results for Stacked Regression models out of all the three regressors used for our study.