The host load prediction problem in cloud computing has also been received much attention. To solve this problem, we have to use the historical load data to predict the future load level. Accurate prediction methods are useful for host load balance and virtual machine migration. Although cloud is likely to grids at some extent, the length of tasks are much shorter and host loads change more frequently with higher noise. The above characteristics introduce challenges for host load prediction. In this paper, based on the proposed exponentially segmented pattern and the corresponding transformation, prediction problem is transformed into the traditional classification problem. This classification problem can be solved based on the traditional methods, and features are given for training the classification model. For achieving accurate prediction, a new feature periodical coefficient is introduced and some existed classification methods are implemented. Experiments on the real world dataset invalidate the efficiency of the new proposed feature, which is in the most effective combinations of features, it increases successful rate(SR) 1.33%~2.82% and decreases the mean square error(MSE) 1.37%~2.91%. And the results also show that support vector machine(SVM) method can achieve nearly the same performance as the Bayes methods and their performance is about 50% higher in successful rate and 17% better in the mean square error compared to the existed methods. The host load prediction problem in cloud computing has also been received useful attention. The above characteristics introduce challenges for host load prediction. In this paper, based on the proposed exponentially segmented pattern and the corresponding transformation, prediction problem is transformed into the traditional classification problem. This classification problem can be solved based on the traditional methods, and features are given for training the classification model. For achieving accurate prediction, a new feature periodical coefficient is introduced and someusing classification methods are implemented. Experiments on the real world datase t invalidate the efficiency of the new proposed feature, which is in the most effective combinations of features, it increases successful rate (SR) 1.33% ~ 2.82% and reduce the mean square error (MSE) 1.37% ~ 2.91%. And the results also show that support vector machine (SVM) method can achieve nearly the same performance as the Bayes methods and their performance is about 50% higher in successful rate and 17% better in the mean square error compared to the existed methods.