Community detection is an unsupervised learning task that discovers groups such that group members share more similarities or interact more frequently among themselves than with people outside groups. In social media, link information can reveal heterogeneous relationships of various strengths, but often can be noisy. Since different sources of data in social media can provide complementary information, e.g., bookmarking and tagging data indicates user interests, frequency of commenting suggests the strength of ties, etc., we propose to integrate social media data of multiple types for improving the performance of community detection. We present a joint optimization framework to integrate multiple data sources for community detection. Empirical evaluation on both synthetic data and real-world social media data shows significant performance improvement of the proposed approach. This work elaborates the need for and challenges of multi-source integration of heterogeneous data types, and provides a principled way of multi-source community detection.