Amazon dataset for ERS-REFMMF

Recommender systems based on matrix factorization act as black-box models and are unable to explain the recommended items. After adding the neighborhood algorithm, the explainability is measured by the user's neighborhood recommendation, but the subjective explicit preference of the target user is ignored. To better combine the latent factors from matrix factorization and the target user's explicit preferences, an explainable recommender system based on reconstructed explanatory factors and multi-modal matrix factorization (ERS-REFMMF) is proposed. ERS-REFMMF is a two-layer model, and the underlying model decomposes the multi-modal scoring matrix to get the rich latent features of the user and the item based on the method of Funk-SVD, in which the multi-modal scoring matrix consists of the original matrix and the preference features and sentiment scores exhibited by users in the reviews corresponding to the ratings. The set of candidate items is obtained based on the latent features, and the explainability is reconstructed based on the subjective preference of the target user and the real recognition level of the neighbors. The upper layer is the multi-objective high-performance recommendation stage, in which the candidate set is optimized by a multi-objective evolutionary algorithm to bring the user a final recommendation list that is accurate, recallable, diverse, and interpretable, in which the accuracy and recall are represented by F1-measure. Experimental results on three real datasets from Amazon show that the proposed model is competitive compared to existing recommendation methods in both stages.