The large number of linked datasets in the Web, and their diversity in terms of schema representation has led to a fragmented dataset landscape. Querying and addressing information needs that span across disparate datasets requires the alignment of such schemas. Majority of schema and ontology alignment approaches focus exclusively on class alignment. Yet, relation alignment has not been fully addressed, and existing approaches fall short on addressing the dynamics of datasets and their size. In this work, we address the problem of relation alignment across disparate linked datasets. Our approach focuses on two main aspects. First, online relation alignment , where we do not require full access, and sample instead for a minimal subset of the data. Thus, we address the main limitation of existing work on dealing with the large scale of linked datasets, and in cases where the datasets provide only query access. Second, we learn supervised machine learning models for which we employ various features or matchers that account for the diversity of linked datasets at the instance level. We perform an experimental evaluation on real-world linked datasets, DBpedia, YAGO, and Freebase. The results show superior performance against state-of-the-art approaches in schema matching, with an average relation alignment accuracy of 84%. In addition, we show that relation alignment can be performed efficiently at scale.