Twitter engagement model for the RecSys 2020 Challenge

Abstract

Recomendation systems is a wide field of research and it is present in many area of our daily life. The RecSys ACM conference is the most important conference in the recommendation area and each year it holds a competition, the RecSys Challenge. The work here presented aims to solve the RecSys 2020 Challenge which consists of giving a certain probability of two Twitter users to interact. We have developed a model which uses the power of Gradient Boosting Trees to combine multiple features we created to represent each interaction between users. Features such as popularity or engagement were combined with and embedding of the tweet text to create an interdisciplinary model that is able to reach 0.75 on the Precision-Recall area under the curve metric and 17.64 on the Relative Cross Entropy. The popularity feature and previous reactions to the same language were discovered as the most relevant features for our model. Regarding the competition, our team reached the ninth place of the challenge.