At a time when social networking sites revolutionize the usages on the Web 2.0, it has become rich, easy, and fun to share private or professional content. Sharing personal information in real-time (such as news, moods, etc…), supports the maintenance of social ties at a high scale. However, the information overload which emerged from the growing quantity of signals exchanged on these services, often in real-time, motivates a regulation of these signals (called mediated interactions), in order to reduce the temporal cost for maintaining social networks, and implied interruptions, which have a negative impact on productivity on tasks that require long-lasting attention. In the frame of this thesis, we are developing a filtering and recommendation system that relies on contextual similarity between users that produce and consume social signals, as relevance criteria.