An algorithm for recommendations that fuses graph representation learning and matrix factorization

Compared to conventional methods of collaborative filtering for recommendations, algorithms that employ matrix factorization are adept at tackling the challenge of sparse data and enhancing the performance of recommendation processes. However, matrix factorization algorithms still have the drawback of not fully utilizing user feature information. To address this, we propose the TransH-MF model. This approach incorporates graph representation learning techniques grounded in matrix factorization. Initially, the TransH graph representation learning algorithm is utilized to project both entities and their relationships into a space, whose vector dimension is lower. Subsequently, the derived vectors are employed to assess the semantic proximity between entities. Ultimately, this proximity data is merged into the process of solving the optimal matrix factorization result. Experiments on the more recent public movielens-latest dataset have proven that the TransH-MF model outperforms the single matrix factorization in regard to both MAE and RMSE metrics.