An Improved PersonalRank Recommendation Algorithm
Jiaqi Liu, Ruilin Dai · 2021
In these years, recommendation systems are broadly researched as well as widely applied in our daily lives. The interaction between users and items is normally expressed by a matrix or graph. Many recommendation algorithms based on graph models have been developed, among which PersonalRank is an algorithm based on the theory of random walk, and NBI is the one based on bipartite network projection. In this paper, we demonstrate the characteristics they share, and fuse them to improve the former. The decay factor in NBI, which is used to inhibit the impact of vertices with high degrees in terms of the initial condition, is applied in PersonalRank in terms of the iteration process. Meanwhile, making use of the idea of the weighted matrix in NBI, we introduce diffusion parameters to generalize PersonalRank further. In addition, the termination condition of PersonalRank is discussed in our work. With the experimental framework shown, we test the algorithms in the benchmark datasets MovieLens. The simulation results show that after our modifications, the precision and coverage of PersonalRank are both significantly improved, outperforming a recent algorithm.