In a multi-label learning system, each sample is associated with multiple class labels at the same time but described by only one feature vector. The common strategy adopted by most existing multi-label classification algorithms is to predict all class labels using the same set of features. It is not the best choice as each label is probably most relevant to its own characteristics. To solve this problem, we propose an improved kNN multi-label classification algorithm with lablespecific features, named IML-kNN. Firstly, the IML-kNN preprocesses the feature vectors of multi-label data and constructs the most discriminative feature for each class of labels. Then, the IML-kNN algorithm is used to do classification based on the obtained characteristics. Experimental results show that the IMLkNN algorithm is obviously superior to the ML-kNN algorithm and other three commonly used multi-label classification algorithms on the yeast and image data sets.