Search and recommendation systems trained directly on query-item pairs suffer from two critical shortcomings: 1) insufficient training data to address the cold start problem (e.g. when a new user comes to the site); and 2) bias towards previously seen pairs, which creates a feedback loop that makes it difficult to generalize to unseen pairs. Wu et al. attempt to overcome these problems through transfer learning of representations taken from item-item prediction tasks. Specifically, they train an encoder that leverages text features of items to predict their neighbors in a correlation graph, and use those encodings to cold-start a recommender system. They find their approach achieved improvements on relevance and user interaction metrics at Google, particularly on short queries that often correspond to broader user intent.