The high pace rise in online as well as offline multimedia unannotated data and associated mining applications have demanded certain efficient mining algorithm. Multiple instance learning (MIL) has emerged as one of the most effective solutions for huge unannotated data mining. Still, it requires enhancement in instance selection to enable optimal mining and classification of huge multimedia data. Considering critical multimedia mining applications, such as medical data processing or content based information retrieval, the instance verification can be of great significance to optimize MIL. With this motivation, in this paper, Multi-Instance, Multi-Cluster based MIL scheme (MIMC-MIL) has been proposed to perform efficient multimedia data mining and classification with huge unannotated data with different features. The proposed system employs softmax approximation techniques with a novel loss factor and inter-instance distance based weight estimation scheme for instance probability substantiation in bags.