The explosive growth of scientific publications and the inaccuracy of metadata cause traditional query-oriented keyword retrieval to hardly meet researchers’ requirements of filtering the desired literature quickly. Graph neural networks (GNNs), which excel at capturing structural graph patterns, have become the practical choice for graph-based recommendation tasks. Although a natural citation network exists for scientific papers, query keywords are isolated nodes and lack connections with the network. Therefore, the related graph convolutional model makes realizing recommendations using keywords in query-oriented scenarios difficult. Furthermore, the time interval is also an essential factor for users to consider, but existing models generally ignore the temporal feature of the publication time of papers. To this end, we propose TextTAGC, a novel paper recommendation algorithm based on temporal-aware graph convolution. The method fuses the graph structure with textual features extracted from paper titles and abstracts through an end-to-end modeling approach. Specifically, this textual information effectively alleviates the cold-start problem of query keywords as isolated nodes in the graph structure. The temporal-aware graph convolutional layer (TAGC) with a prediction layer learns the temporal patterns embedded in citation relations during aggregation. Experimental results show that the proposed model improves MAP metrics by about 27 and 21% compared to the baseline approach in query-oriented cold-start scenarios on AAN and DBLP datasets.
Liu et al. (Sat,) studied this question.