A machine learning-based recommendation model for bipartite networks

Abstract

Listen

Translate article

Online user reviews on a product, service or content has been widely used for recommender systems with the spread of the internet and online applications. Link prediction is one of the popular recommender system approaches. It can benefit the structure of a social network by mapping item reviews of users to a bipartite user–item graph structure. This study aims to investigate how topological information, namely neighbor-based, path-based and random walk-based network similarity metrics, improve the prediction capability of a recommendation model. This study proposes a supervised machine learning-based link prediction model for weighted and bipartite social networks. The input features of the machine learning model are extended versions of similarity metrics for weighted and bipartite networks. Our proposed model provides 0.93 and 0.9 AUC values for the Goodreads and MovieLens datasets, respectively. Random forest and extreme gradient boosting as the ensemble models achieved the highest performances for ItemRank metric in both datasets.