Complex networks have been widely used to model complex systems composed of interacting entities. It is possible to recommend new relationships between entities according to the network topology and the entities’ properties. The likelihood of forming a missing or potential relationship is often captured by similarity measures. As community clusters are also based on the entities’ similarities, traditional link recommendation/prediction methods naturally tend to recommend links within a community. However, potential links valuable across communities are often overlooked and may cause the problem of information cocoons. We focus on link recommendation across communities based on homogeneous and heterogeneous information networks, which aims to improve the diversity of recommender systems. We propose to solve this problem using a novel similarity calculation and heterogeneous network embedding methods. Our comprehensive experiments on real-world datasets and synthetic datasets show that our methods strike a good balance between accuracy and efficiency, while generating valuable unconventional recommendations in practical application scenarios.