In order to alleviate the problem of data sparsity in recommendation systems, based on the advantage of convolutional neural networks (CNNs) in capturing local features, we propose a local similarity prediction recommendation model based on improved CNN (LSPCNN) by adding an adjustment layer. The new model iteratively adjusts the initial user-item scoring matrix to localize user’s interest preference. And then CNNs are integrated to predict the missing score and achieve personalized recommendation. Experimental results show that the MAE value of the LSPCNN model under different degrees of data sparsity decreases by an average of 4% compared with the traditional recommendation methods, and it effectively alleviates the data sparsity problem and improves the performance of recommendation systems.