Text Embeddings for Recommender Systems)

Article and social media recommender systems aim to enrich a user experience of browsing the web and help to find something genuinely interesting from an infinite space of available options. The goal of this project is to incorporate deep learning methods to predict interesting articles for a user on the real-life data from production logs in Yandex.Zen. In particular, this work is targeted to leverage varied contextual information of articles, such as text, titles, URLs and images to improve the quality of recommendations. The contribution of this work is twofold. First, we proposed a neural network approach to learn low-dimensional document representations in an unsupervised fashion. Obtained embeddings have a good property, such as similar documents have small cosine distance in the latent space. Second, we proposed a model that learns relevance metric between a user and a document, given the history of the user's previous interactions with the service. The proposed methods draw on the attention module to aggregate dynamically representative articles from a user’s browsing history in respect to a candidate article. We evaluated the method both as a standalone model and as factors for the current production model, based on the gradient boosting over decision trees. The experiments have shown that our model significantly improved the accuracy of recommendations on historical data.