Decentralised social networks promise to deliver highly personalised, privacy-preserving, scalable and robust implementations of key social network features, such as search, query extensions, and recommendations. Such systems go beyond traditional online social networks by leveraging implicit social ties to implement personalised services. Yet, current decentralised social systems usually treat all users uniformly, when different sub-communities of users might in fact work best with different mechanisms. In this paper, we look at the specific case of decentralised social networks seeking to cluster users exhibiting similar behaviours to provide decentralised recommendations. These decentralised recommendation systems typically rely on a single metric applied uniformly to all users to extract similarities, while it seems natural that there is no such one-size-fits-all approach. More specifically we show in this paper, using a real Twitter trace, that (i) individual users can benefit from a personalised strategy in the context of decentralised recommendation systems, and that (ii) overall system performance is improved when the system accounts for the varying needs of its users i.e. when each user is allowed to diverge and use its optimal strategy.