Social networks are known to be assortativewith respect to many attributes, such as age, weight,wealth, level of education, ethnicity and gender: simi-lar people according to these attributes tend to be moreconnected. This can be explained by influences and ho-mophilies. Independently of its origin, this assortativitygives us information about each node given its neigh-bors. Assortativity can thus be used to improve indi-vidual predictions in a broad range of situations, whendata are missing or inaccurate. This paper presentsa general framework based on probabilistic graphicalmodels to exploit social network structures for improv-ing individual predictions of node attributes. Using thisframework, we quantify the assortativity range lead-ing to an accuracy gain in several situations, with vari-ous individual prediction profiles. We finally show howspecific characteristics of the network can enhance per-formances further. For instance, the gender assortativ-ity in real-world mobile phone data drastically changesaccording to some communication attributes. In thiscase, using the network topology indeed improves localpredictions of node labels, and moreover enables infer-ring missing node labels based on a subset of knownvertices. In both cases, the performances of the pro-posed method are statistically significantly superior tothe ones achieved by state-of-the-art label propagationand feature-extraction schemes in most settings.