Prediction of links in networks often requires an effort in feature engineering and utilization of expert knowledge. Much of today's research are primarily domain-specific, simply tuning variations of basic methods to fit a particular purpose.

Network embedding is a method to learn low-dimensional representations of vertices in a network, and recently such methods have been significantly improved by borrowing from advancements in the field of natural language processing.

We propose to apply feature learning on topological network data and thereby learn network representations specifically for link prediction in bipartite networks. We present a novel algorithmic framework for exploiting recent methods of network embedding onto bipartite networks with the purpose of predicting links. Applications of our framework may be in any field having bipartite network representations.

We demonstrate improvements over state-of-the-art techniques for link prediction. Besides achieving better precision, our proposed framework has further advantages including high modularity, scalability and a by-product in the form of vertex embeddings, useful for other unrelated machine learning tasks.