Data mining and machine learning methods have been playing an important role in searching and retrieving multimedia information from all kinds of multimedia repositories. Although some of these methods have been proven to be useful, it is still an interesting and active research area to effectively and efficiently retrieve multimedia information under difficult scenarios, i.e., detecting rare events or learning from imbalanced datasets. In this paper, we propose a novel subspace modeling framework that is able to effectively retrieve semantic concepts from highly imbalanced datasets. The proposed framework builds positive subspace models on a set of positive training sets, each of which is generated by a Gaussian Mixture Model (GMM) that partitions the data instances of a target concept (i.e., the original positive set of the target concept) into several subsets and later merges each subset with the original positive data instances. Afterwards, a joint-scoring method is proposed to fuse the final ranking scores from all such positive subspace models and the negative subspace model. Experimental results evaluated on a public-available benchmark dataset show that the proposed subspace modeling framework is able to outperform peer methods commonly used for semantic concept retrieval.