Owing to the rapid proliferation of Internet service technologies, the development of social network analysis is ever-increasingly important in recent years. As large datasets in social networks becomes available, recommendation plays a more and more important role in our daily lives. Recommendation approaches automatically prune large information to recommend the most relevant data to users by considering their preferences. Recent studies demonstrate that the efficiency of social networks could be exploited by improving the performance of recommendations. In this article, a novel recommendation approach is proposed to effectively extract dense subsets from sparse data set of micro blog social network, and cluster the whole user group into categories based on content similarity to produce better recommendation results. Through groups of reasonable experiment implementations with real data crawled from micro blog social network, the performances of this new proposed approach and other classical existing recommendation approaches are evaluated and compared by various measurable parameters. The experimental results demonstrate that the proposed approach could greatly improve the recommendation accuracy rate, recall rate and comprehensive measurable indexes when compared with other studied recommender algorithms. On the other hand, the computation overhead of the proposed approach is smaller than that of the other ones.