Abstract: In order to improve the efficiency of phishing detection, a new algorithm was proposed to improve the traditional random forest algorithm. Potential association rules between web features are mined and used to partition the data set, in order to distinguish the features of different structures and calculate the weight of different data space to determine the scale of the selection. After selection of data, training data sets need to be aggregated and clipped to optimize the establishment of forests. Websites are trained and predicted using voting in decision forest. Experiments result shows that the new algorithm has obvious advantages in efficiency and effectiveness compared with the other two algorithm.