Non-negative matrix factorization belongs to the group of linear dimensionality reduction methods and its main goal is to approximate non-negative matrix with the product of two low-rank non-negative matrices. This kind of transformation identifies latent features preserving non-negative structure of the original data which leads to easier interpretability. The most widely used cost function is based on Frobenius norm, while Kullback-Leibler divergence has shown to be effective for sparseness. Taking into account that the solution to this problem is not unique, and in order to decrease approximation error, it is essential to estimate reduced rank, i.e. lower dimension into which the original data is being transformed. One of the key factors here is the number of classes recognized by the algorithm. Both reduced rank estimation and approximation factors are typically obtained using algorithms based on alternating least squares, and, more often, multiplicative update thanks to its simplicity. Due to its versatile applications such as topic modeling, audio source separation, clustering and temporal segmentation, non-negative matrix factorization found its way into various fields characterized by non-negative data, such as bioinformatics, astronomy, music, textual analysis, etc. One of the recent applications is regarding link prediction in networks, where new friendships, coauthorships and even neural connections can be successfully obtained using perturbations or attribute matrices. Using the coauthorship network CROSBI as an example, in this work it was shown that non-negative matrix factorization in combination with perturbations and attribute matrices based on the network topology, such as shortest path distance and sum of neighbors, outperforms results obtained by classical link prediction methods