As a particular type of social networking site, online academic communities have revolutionized the way researchers collaborate and communicate with each other. Accompanying the growth in the number of users registered on online academic communities, the information overload problem presents a great challenge for researchers trying to find relevant and reliable friends there. Different from friends in conventional social networking sites, friends in online academic communities are denoted as scholar-friends in this research. Scholar-friend recommendation in online academic communities involves different entities (e.g., researchers, research articles, affiliations, research interests, and status updates) and various relationships among entities (e.g., scholar-friend relationship of researchers, post relationship between researchers and status updates, and writing relationship between researchers and papers), which constitute a complex heterogeneous network. By leveraging the entity and relationship data in online academic communities, this research proposes a heterogeneous network-based approach to recommending scholar-friends where information gain is used to identify valuable meta paths and a regularization-based optimization is employed to make personalized recommendations for each individual researcher. Experimental results based on historical data and user experiments demonstrate that the proposed approach can achieve better performance compared to some baseline approaches. Additionally, we further discuss how the meta paths and corresponding learned weights can help to understand researchers' preferences and behaviors.