Leveraging User Embeddings and Text to Improve CTR Predictions With Deep Recommender Systems
Carlos Miguel Patiño, Camilo Velásquez, Juan Manuel Muñoz, and
3 more authors
In Proceedings of the Recommender Systems Challenge 2020, 2020
Predicting user engagement, often framed as a CTR prediction problem, is important to maximize user satisfaction in social networks. The 2020 Recsys Challenge was sponsored by Twitter and set the goal of predicting four types of user engagement using a dataset with 160 million tweets. Our approach extracted information from the tweet’s text tokens and built optimized user embeddings. We designed our model based on ideas from recommender systems and deep learning that had been successful in CTR prediction tasks. We show that our modifications to existing state-of-the-art architectures and feature engineering improved the model’s ability to predict user engagement. Factored’s team was called Los Trinadores and had the 6th best submission of the challenge with an overall score of 22. The code for our solution is available at https://github.com/factoredai/recsys20-challenge/.