In recent decades, with the advancement of information and communication technology, the need for organizing and optimizing informational and communicational processes among individuals and various resources has gained increased importance. One effective approach in this field is the use of recommender systems. These systems can significantly enhance user experience either offline, by preparing a list of the best recommendations, or online, in real-time. One of the challenges of recommender systems is achieving a balance between the speed of delivering recommendations and their quality. This paper presents the development and evaluation of a recommender algorithm named "Newpixie," which aims to provide recommendations based on common interests among users. The Newpixie algorithm utilizes a scalable, bipartite graph comprising pins and boards, where each board contains a number of related pins on a specific topic. In the offline phase of the algorithm, strong associations between pins are extracted using association rules in data mining and applied as virtual edges in the graph. In the online phase, biased random walks on the strong links between pairs of items within the boards are employed to generate suitable recommendations for users. The performance of the Newpixie algorithm is evaluated against the baseline pixie method based on three evaluation metrics. Experimental results indicate that the Newpixie algorithm improves recommendation speed by approximately 17%, recommendation quality by about 34%, and the recall@k metric in link prediction by roughly 20% compared to the baseline pixie method.