Journal of Digital Information Management H. BAI, G. YU, XY. TIAN, 14 2 2016 Weibo is an extensively used social network tool in China and has become a popular platform for disaster information management. This popular microblogging service offers massive firsthand information regarding the state and emotions of victims in a disaster situation. Identifying negative sentiment messages from the large-scale and noisy Weibo stream is a fundamental and challenging undertaking. Therefore, based on the characteristics of negative Weibo messages concerning disaster events, a novel feature selection algorithm called combined frequent pattern (FP)-growth and mutual information theory (CFM) algorithm, was proposed to improve the traditional machine learning approaches in this study. The CFM algorithm mined two FPs via FP-tree, and the mutual information between two frequent items was calculated to determine the most frequent and tight features for negative-sentiment Weibo messages detection. After that, the experimental analysis was conducted to test the proposed novel feature selection algorithm and to explore a suitable sentiment classifier for disaster-related Weibo messages. The analysis employed actual disaster-related Weibo message data set, which included 2,913 negative messages and 2,913 un-negative messages. Results demonstrate that the CFM algorithm performs well in the feature selection process. In particular, this algorithm exhibits the best performance in the support vector machine classifier with 89.34% accuracy. Therefore, the CFM algorithm is an efficient feature selection algorithm for negative-message classification in a post-disaster situation. This algorithm also offers a novel method to reduce the feature dimension in other text classification areas.