In the matrix factorization (MF) based collaborative filtering recommendation method, the most critical part is to deal with the interaction between the features of users and items. The mainstream approach is to use the inner product for MF to describe the user-item relationship. However, as a shallow model, MF has its limitations in describing the relationship between data. In addition, when the size of the data is large, the performance of MF is often poor due to data sparsity and noise. This paper presents a model called PIDC, short for potential interaction data clustering based deep learning recommendation. First, it uses classifiers to filter and cluster recommended items to solve the problem of sparse training data. Second, it combines MF and multi-layer perceptron (MLP) to optimize the prediction effect, and the limitation of inner product on the model expression ability is eliminated. The proposed model PIDC is tested on two datasets. The experimental results show that compared with the existing benchmark algorithm, the model improved the recommendation effect.