Taboola’s content discovery platform aims to perform ”reverse search”, i.e match content to users who are likely to engage with it. Taboola’s content recommendations serve billions of recommendation per day, with a user base of hundreds of millions of active users. Our goal is to train computational models in order to predict the probability of an ad being clicked by a user in a specific content.
Deep Density Network (DDN) was introduced, a unified deep neural network model which predicts click-through-rate (CTR) of an ad in a specific context incorporating both measurement and data uncertainty, and having the ability to be trained end-to-end while facilitating the exploitation/exploration selection strategy. DDN employs state-of-the-art Computer Vision and Natural Language Processing technologies and exploiting deep learning ability to capture non-linearities has enabled us model complex ad-context relations and integrate higher level representations of data sources such as contextual, textual and visual input.
DDN runs in production serving billions of recommendation per day to hundreds of millions of active users. Furthermore, outperforms by far previously traditional linear models, and while employing it in a multi-arm bandit setting, yields an adaptive selection strategy that balances exploitation and exploration and maximizes the long term reward.
Yoel Zeldes, Stavros Theodorakis, Efrat Solodnik, Aviv Rotman, Gil Chamiel, Dan Friedman, “Deep density networks and uncertainty in recommender systems”, https://arxiv.org/abs/1711.02487