The New York Times allows readers to identify their interests to provide them with more personalized story recommendations. Historically, interest-based recommendations drew on tags specified by authors and other experts. In this post, former PTK author Joyce Xu describes an approach wherein her team developed multi-label classification models that predict interests based on article text using noisy labels (e.g., imperfectly specified tags). She describes the model architecture (including methods for embedding vectors and the classification task) but also discusses how and why NYT plans to implement a human (editor)-in-the-loop approach. Most importantly, #bringbackJoyce!