Link Prediction for Recommendation based on Complex Representation of Items Similarities

Recommender systems have different goals, including coverage and hit rate. A graph-based recommender system that includes users' ratings of items can be represented as a bipartite graph, where users and items are nodes and ratings are treated as edges. The recommendation task in bipartite recommender systems can be modeled as a subproblem of link prediction. Modified link prediction methods are used in related research to distinguish between fundamental relational dichotomies like versus dislike, and similar versus dissimilar. However, similarities between users and items should be noticed. In this paper, we present a novel method that focuses on developing an enhanced adjacency matrix designed to bolster link prediction and recommendation accuracy. This innovative approach strengthens our ability to capture intricate user behaviors. The results over a real-world dataset indicate the effectiveness of the proposed approach.