An Improved Recommendation Algorithm For Polarized Population

The performance of various recommendation algorithms is getting stronger and stronger, which also brings new social problems and group polarization. The recommendation algorithm creates a user’s selective access to information, and the continuous iteration of the algorithm and the continuous enhancement of selective information access also lead to the echo chamber effect. To re-apply the collaborative filtering algorithm in polarized crowds and eliminate its echo chamber effect, a collaborative filtering algorithm adapted to polarized crowds is proposed in this paper. The algorithm can determine whether the user is a polarized crowd by using user characteristic information and user behavior information and make collaborative filtering that adapts to the polarized crowd. Specifically, the concept of “polarization fraction” proposed in this paper is a specific score calculated by weighting user characteristic information and user behavior information, and the score is compared with a given threshold to determine whether the user is a polarized crowd. For the polarized population, the user characteristic information and user behavior information are processed by word2vec to get the dangerous word and converted the data from floating-point values to dense vectors and k-means clustering them. After clustering, calculate the distance between the cluster center and the dangerous word vector, which is less than the given threshold, and delete the cluster. For evaluation, two real data-sets ml-latest and ml-1m were tested. The experimental results show that the algorithm can better adapt to the polarized population and has good performance.