The work of 21 integrated convolutional neural network (CNN) into PMF, to capture contextual information of item description documents to further enhance rating prediction accuracy. Their model specifically assumed that latent model for an item is not generated from the latent models distribution as in the PMF model, instead is generated from three variables; 1) internal weights of the CCN; 2) item representation from the document description; and 3) epsilon which is the Gaussian noise. In other words, instead of using Gaussian prior with zero mean on items, they defined the conditional distribution of items to be given by the Gaussian variable (set of item description documents) with mean a document latent vector obtained from the CNN model and the Gaussian noise was used as variance, hence integrating CNN and PMF.