Abstract
As an effective tool of information filtering, the network-based recommendation algorithms encounter the challenging problem of recommendation bias induced by the object heterogeneity. Previous solutions usually make the improvement based on some specific algorithm, however, are difficult to generalize to different algorithms. In this article, we propose an improved model with a general formula, by inhibiting recommendation bias described by the eigenvalue and eigenvectors of the algorithm similarity matrix, and applied the model into ten different algorithms. Based on four real recommender systems, the experimental results show that nearly all the algorithms are improved in three aspects of recommendation accuracy, diversity and novelty, for all the four datasets. The recommendation accuracy of cold objects is also elevated. Especially, two excellent algorithms are further improved without introducing any other parameter. Our work may shed a new light on developing general recommendation algorithms from the perspective of revealing intrinsic feature in recommender systems. (C) 2022 Elsevier B.V. All rights reserved.
NSTL
Elsevier