isprec++: learning edge type importance in network-oriented paper recommendation

With the spread of the internet and the possibility of online access to articles, a wide range of scientific articles are available to researchers, while finding relevant articles among this substantial number of articles turns out to be a real dilemma. to solve this problem, several scientific paper recommendation algorithms have been proposed. most of these algorithms suffer from some drawbacks that limit their usability. for example, many of these recommendation methods are designed to recommend papers only to users who had published articles before and can’t support new researchers. also, they usually do not utilize many important features of articles each of which can have a role in determining the relevance of the articles to users. to address these concerns, in this paper, we present the novel method of integrated scientific paper recommendation with an edge-weight learning approach, called isprec++, as an extended version of isprec that focuses on learning the weights of edge types in heterogeneous information networks based on users’ preferences. isprec++ sets the weights of edges in spin using a bayesian personalized ranking (bpr) based method and utilizes gradient descent to optimize its objective function. thereafter, it exploits a limited random-walk algorithm for a top-n recommendation. extensive experiments on a real-world dataset demonstrate the significant performance superiority of isprec++ compared to the state-of-the-art scientific paper recommendation algorithms.