Online social networks support users in a wide range of activities, such as sharing information and making recommendations. In this talk, I will talk about a human-generated recommendations of friendship based on hashtags that emerged in Twitter around 2010. I will present a study of a large-scale corpus of human friendship recommendations, using a large corpus of tweets gathered over a 24 week period and involving a set of nearly 6 million users. I will show that these explicit recommendations had a measurable effect on the process of link creation, increasing the chance of link creation between two and three times on average, compared with a recommendation-free scenario. Also, ties created after such recommendations have up to 6% more longevity than other Twitter ties. Finally, I will talk about a supervised system that ranks our user-generated recommendations, surfacing the most valuable ones with high precision (0.52 MAP). We find that features describing users and the relationships between them are discriminative for this task. After the talk, we will carry out some examples on the collection of online data