Trust-Distrust Aware Recommendation by Integrating Metric Learning with Matrix Factorization

With the upsurge of e-commence and social activities on the Web, Recommender system has attracted widespread attention from both researchers and practitioners. Motivated by the fact that trusted users tend to have small distance whereas distrusted users tend to have large distance, we propose to model trust aware recommendation based on distance metric learning. Furthermore, by incorporating with classical matrix factorization method, we build an integrated optimization framework and convex loss function. Gradient descent method is employed to optimize the loss function. Experiments are conducted on the Epinions dataset, which shows that the performance of the proposed method is remarkably superior to competitive methods in terms of precision, recall and F1-measure, and compatible in terms of MEA and RMSE, demonstrating advantages of modeling recommendation with trust-distrust aware metric learning and matrix factorization. To the best of our knowledge, this is the first attempt in modeling and optimizing recommendation in a unified framework of trust aware distance metric learning and matrix factorization.