Recommender systems are one of the fields of information filtering systems that have attracted great research interest during the past several decades and have been utilized in a large variety of applications, from commercial e-shops to social networks and product review sites. Since the applicability of these applications is constantly increasing, the size of the graphs that represent their users and support their functionality increases too. Over the last several years, different approaches have been proposed to deal with the problem of scalability of recommender systems’ algorithms, especially of the group of Collaborative Filtering (CF) algorithms. This article studies the problem of CF algorithms’ parallelization under the prism of graph sparsity, and proposes solutions that may improve the prediction performance of parallel implementations without strongly affecting their time efficiency. We evaluated the proposed approach on a bipartite product-rating network using an implementation on Apache Spark.