Abstract: There is a great potential for interest modeling on description and extension as the personalization services enrich the portal. A novel personalization interest model based on latent interest semantic description(PIM-LISD) was proposed. As a finite mixture, it was represented with personalized interests elicited implicitly on portal and latent interest relating semantic descriptions. The initialization method was to fit the posterior probabilities along with the different reasonable prior predictions, so the interpretative and adaptive capacity of the model could be perfect. During the building procedure, an improved tempered expectation maximization (TEM) was used for variational expectation maximization estimation. The experiments indicate that the proposed modeling method can not only avoid the complication of potential expensive modeling-building stage effectively, but also increase the prediction precision, therefore proving its validity and feasibility.