Abstract: With the development of information technology, people can get more and more information nowadays. To help users find the information that meets their needs or interest among large amount of data, personalized recommendation technology has emerged and flourished. As a most widely used and successful recommendation technique, collaborative filtering algorithm has widely spread and concerned many researchers. Traditional collaborative filtering algorithms face data sparseness and cold start problems. As traditional algorithms only consider the limited data, it is difficult to estimate the accurate similarity between users, as well as the final recommendation results. This paper presents a kernel-density-estimation-based user interest model, and based on this model, a user-based collaborative recommendation algorithm based on kernel method is proposed. Through mining users' latent interest suggested by the limited ratings, the algorithm can well estimate the distribution of users' interest in the item space, and provide a better user similarity calculation method. A distance measurement based on classification similarity is proposed for the kernel methods, and two kernel functions are investigated to estimate the distribution of user interest. KL divergence is utilized to measure the similarity of users' interest distribution. Experiments show that the algorithm can effectively improve the performance of the recommendation system, especially in the case of sparse data.