The continuous and progressive growth of the need for knowledge extraction from continuous data streams, in an exponential way, has favored the emergence of a new research axis from the semantic web community. In the few last years, many semantic data stream processing systems have been proposed by combining Data Stream Management Systems (DSMS) technologies and Semantic Web technologies (RDF1/SPARQL2) for annotation, publication and reasoning on these data streams. However, considering their infinite volume and unknown velocity, processing and storing their contents remain impossible, which leads to introduce techniques for reducing load and/or summarizing data. In this context, we propose a graph-oriented approach to reduce the semantic data streams volume. In order to validate our approach, we implemented it using Simple Random Sampling and Stratified Random Sampling and we experimented it using the CSRBench benchmark. Our approach allows to maintain the data consistency and their semantic level.