A scientific paper recommendation method using the time decay heterogeneous graph

Finding appropriate and relevant papers about a project in various digital libraries with millions of scientific papers is challenging for researchers, resulting in a research innovation gap because of incomplete literature retrieval. A query-oriented paper recommendation (QPR) is a feasible way to improve the efficiency of literature retrieval in scientific research, and the graph-based method is one of the best solutions for QPR. However, current graph-based QPR methods still have the defeats of low precision and over-weighting. This paper proposes a query-oriented paper recommendation method using the Time Decay Heterogeneous Graph (TDHG) to improve the recommendation quality. TDHG is a four-layer heterogeneous graph combing the time decay characteristics in academic literature. We also used author rank to highlight contributions, the Author-Topic model to extend relations in heterogeneous graphs, and the Random Walk with Restart algorithm to rank papers. We designed three time-decay vectors and compared their impact on overcoming over-weighting caused by applying Random Walk with Restart algorithm to the original heterogeneous graph. Our experiments show linear time-decay vectors cannot balance the importance and timeliness of academic papers, while log time-decay vectors and sqrt time-decay vectors effectively solve the over-weighting problem. The experimental results show that the time-decay vector brings about an 11% and 8% improvement in Mean Average Precision (MAP) on the AAN and DBLP datasets, respectively.