KDE–CSSD, an tree structure algorithm was proposed for the prediction of protein function based on class hierarchy to solve the issues of high computational cost on label classes through direct learning classification model and of train data skew on class hierarchy among middle or lower level nodes. The algorithm firstly projected label vector onto principle components of label kernel by means of learning less regression models, then, the predicted numeric vector were back projected onto their original vector space, finally, the predicted 0 or 1 label vector meeting tree hierarchy constraint were obtained using compressed sort and selection algorithm. The experiments, adopted precise rate and recall rate as criterion on 12 genomic benchmark data sets, proved that the KDE–CSSA algorithm outperformed the outstanding CLUS–HMC algorithm.