Abstract: In problem transaction methods to solve the problem of multi－label classification，binary relevance is a commonly used method. However，there usually exits an assumption that labels are independent of each other in binary relevance，which will lead to the neglection of the correlations between labels. Classifier chain for multi－label classification takes label correlations into account through passing label information between classifiers，which is improved based on binary relevance. In this way，this improved method overcomes the label independence in binary relevance. However，the order of classifier chain is arbitrary in this algorithm. Moreover，different order of the classifier chain will lead to different classification accuracy of the learning task. In kernel theory，kernel alignment is a quantity to measure the degree of consistency between two kernels. In order to solve the problem of arbitrary classifier chain order，kernel alignment is introduced into classifier chain for multi－label classification to sort the classifier chain，and two algorithms based on kernel alignment are proposed. The first algorithm maximizes the kernel alignment value between kernel function in feature space and convex combination of ideal kernels in label space，where each ideal kernel is defined by each label in label space. Then the weight in the convex combined ideal kernel is obtained through solving this optimization problem，and the classifier chain is sorted based on the weights of each ideal kernel. The second algorithm directly computes the kernel alignment values between kernel in feature space and each ideal kernel in label space，thus the alignment values corresponding to each label are obtained. Then the order of classifier chain can be sorted based on these alignment values. Our idea of introducing kernel alignment into classifier chain for multi－label classification to order the classifier chain is experimentally demonstrated to be effective.