With the current rapid development of mobile Internet, the information overload problem that people face is particularly serious, which makes it a big challenge to do particular users' personalized recommendation in the big data scenario. In order to further improve the timeliness, accuracy of recommendation and ease the problem led by large amount of data, we propose a optimized matrix decomposition recommendation algorithm under the environment of big data in this paper. This algorithm integrates users and the similarity computation of items on the basis of the traditional matrix decomposition algorithm. In the process of training objective function, we enhance the recommendation accuracy by taking in account of users and k nearest neighbors' similarity computation of items. Taking Spark's advantage on memory computing and iterative computing, we design an algorithm using matrix decomposition and nearest neighbor fusion under the Spark framework. Experiments conducted on the classical MovieLens dataset show that our proposed algorithm can deal with data sparseness well, improve recommendation accuracy to some extent, and has a better computational efficiency in the comparison with traditional matrix decomposition recommendation algorithms.