Recommender systems are facing cold start problems usually referred to as the item cold start and the user cold start. To solve both (1) the item cold-start problem and (2) the problem of recommending to the weakly engaged user, I introduce my contribution to the topic, a novel framework based on non-negative matrix factorization (NMF). The framework learns collective representations from the item’s features and the users’ feedback. Experiments on news articles and emails show the effectiveness of the method. Afterwards, in order to solve the problem of recommending to the weakly-engaged user I introduce two novel techniques: (1) By relying on endogenous information, I show how our framework improves the recommendations of news articles for the weakly-engaged user, this is around 75% of users (i.e. large coverage); (2) By relying on exogenous information extracted from search queries submitted on Yahoo, I illustrates how the recommendations of news articles can be improved for 200K users.