Microblogging filtering is intended to filter out irrelevant content, and select useful, new, and timely content from microblogs. However, microblogging filtering suffers from the problem of insufficient samples which renders the probabilistic models unreliable. To mitigate this problem, a novel method is proposed in this study. It is believed that an explicit brief query is only an abstract of the user's information needs, and its difficult to infer users' actual searching intents and interests. Based on this belief, a filtering model is built where the multi-sources query expansion in microblogging filtering is exploited and expanded query is submitted as users interest. To manage the external expansion risk, a user filter graph inference method is proposed, which is characterized by combination of external multi-sources information, and a risk minimization filtering model is introduced to achieve the best reasoning through the multi-sources expansion. A series of experiments are conducted to evaluate the effectiveness of proposed framework on an annotated tweets corpus. The results of these experiments show that our method is effective in tweets retrieval as compared with the baseline standards.