Fair Personalized Recommendation through Improved Matrix Factorization by Neural Networks.

Existing recommendation systems mainly focus on the improvement of accuracy but neglect to involve non-popular items, which leads to the unfairness problem of recommendation results. More specifically, the low exposure rate of non-popular items impairs the diversity of recommendation results and thus weakens the customer experience for online e-commerce platforms. To address this problem, we propose a Fair Personalized Recommendation Algorithm(FPRA) combining improved Matrix Factorization and Back-Propagation Neural Network, which can simultaneously acquire both the accuracy and the fairness of recommendation results. In the FPRA, popular items and non-popular items are separated according to a pre-defined proportion. On the one hand, we consider that the influence of scenario factors (i.e., browsing location, clicking number, etc.) will generate users’ assessment bias for popular items, and we leverage BP Neural Networks to train the bias when recommending popular items. On the other hand, we design a new function, namely Fairness & Accuracy function, to achieve the goal that non-popular items will be presented to customers more frequently. Using this function, the goal can be achieved by searching for the optimal solution under the condition constraints. Experimental results show that the effectiveness of our proposed that FPRA greatly outperforms other baseline methods.