After grouping users using divisive hierarchical clustering for each hour, users with the same application preference will be classified into the same “interest cluster” (there are 11 interest clusters in each hour, a user can only be classify into one interest cluster). Then the normalized entropy values are calculated for each cluster. The bigger normalized entropy value is, the more application categories the user visits in one hour, e.g. the user tends to visit more categories of applications in one hour. In contrary, a small normalized entropy value implies that user’s interest is limited. Namely, users in an interest cluster which has small normalized entropy value tend to concentrate on very few application categories. Fig. 9 shows the normalized entropy of each interest cluster for every hour in the week.