Accurate link prediction across a large user base has become a challenging problem as online social networking platforms alter the ways and means of communication. Numerous applications, including friend recommendations, news commentary, and product recommendations, are affected by the issue. In this research, we provide a brand-new algorithm to address this issue. Due to their limitations in making full use of information or capturing all the features, the present online social network link prediction algorithms have various shortcomings in their link prediction accuracy. This study presents a novel formulation of the link prediction issue as a matrix denoising problem. We first suggest and thoroughly describe an unsupervised marginalized denoising model (USMDM). A mapping function that can find patterns in a massive amount of user data and comprehends the topological structure of social networks is the basis of the USMDM. A target matrix is projected onto the observed matrix via the mapping function. The initial matrix in the learning process is replaced with a low-rank matrix to increase effectiveness and avoid overfitting. The function can be trained on small datasets using the weak law of large numbers. Experiments are carried out on four actual social networks to show how well the suggested algorithm works, and the outcomes show how well the model works