Trust is a very significant notion in social life, and even more in online social networks where people from different cultures and backgrounds interact. Weighted Signed Networks (WSNs) are an elegant representation of social networks, since they are able to encode both positive and negative relations, thus allow to express trust and distrust as we know them in the real world. While many trust inference algorithms exist for traditional unsigned networks, distrust makes it hard to adapt them to WSNs. In this paper, we propose a new unsupervised trust inference algorithm based on collaborative filtering (CF), where we consider the trustors as users, the trustees as items, and agreement as a local similarity metric to predict trust values in signed, and unsigned, networks. In addition to its prediction performances, experiments on four real-world datasets show that our algorithm is very robust to network sparsity.