In video data， one concept in one shot are usually dependent on others concepts. Several semantic concepts appearing in one time often determine the presence of other concepts. An inter-concepts association and dependency multi-label video semantic concept classification approach is proposed in this paper. In order to generate association and dependency relation between concepts， join and prune phases are used to extract potential itemsets. After calculating the minimum support of each itemset， frequency itemsets meeting the user specified minimum support are selected. In the iteration process of generation frequency itemsets， compound labels with strong association and dependency relation of inter-concepts are obtained. Finally， compound labels are considered as a single label in the annotation step. Experiments on real-world multi-label media data show that this method the methods beat accuracy of existing multi-label learning methods with statistically significant improvements.