Recommendation systems represent personalized services that aim at predicting users' interest on information items available in the application domain. The computation of the neighbor set of users or resources is the most important step of the personalized recommendation system, and the key to this step is the calculation of similarity. This paper analyzes three main similarity algorithms and finds deficiencies in these algorithms, which affect the quality of the recommendation system. Then the paper proposes a new similarity algorithm Simi-Huang, which effectively overcomes the above-mentioned drawbacks. Experiments show that Simi-Huang algorithm is better than the three main similarity algorithms in the computation of accuracy, especially when the data is sparse. Under different training models, Simi-Huang is best in accuracy of all the algorithms; the smaller the training model is, the more accurate the algorithm is.