With the growing adoption of Linked Open Data (LOD), many distributed knowledge bases have been developed using the RDF format. These knowledge bases each have unique characteristics, and querying across them can reveal richer and more comprehensive insights. To support such cross-repository querying, federated RDF query processing is essential. However, the performance of federated query execution largely depends on efficient source selection—the process of identifying which data sources are relevant to a given query. In this paper, we propose a novel source selection method for federated RDF queries based on a new summarization technique. Our approach precomputes the existence of shared elements between predicate pairs within each data source, considering three types of join patterns: subject–subject, subject–object, and object–object. These relationships are stored in a matrix-form summary for each data source. During query execution, our method identifies predicate pairs from triple patterns that share join keys within a basic graph pattern (BGP) and uses the precomputed summaries to efficiently determine the relevant data sources. Experimental results on the FedBench benchmark show that our method improves the efficiency of source selection and significantly reduces overall query execution time.