Over the past few decades, the exponential growth of digital publishing has granted researchers access to an ever-expanding corpus of scientific articles. However, the resulting topical diversity makes it increasingly difficult to locate truly relevant works. Existing citation recommendation models seek to personalize suggestions but often (i) under-exploit the rich, multi-relational structure of heterogeneous citation networks, (ii) inadequately model semantic interactions among diverse entities, and (iii) lack mechanisms to highlight the most influential factors or to provide transparent explanations. To address these gaps, we propose CR-KGEB, an encoder–decoder framework that integrates RotatE knowledge-graph embeddings with SPECTER content representations and a BiLSTM attention module over a k -partite graph of authors, papers, tags, and publication times. By jointly modeling an author’s published articles, citation history, and candidate papers–and dynamically weighting salient signals–CR-KGEB delivers both highly accurate and interpretable recommendations. On the DBLP-V12 and V13 datasets, CR-KGEB increases MAP from 0.592 to 0.625 and recall@100 from 0.804 to 0.829 on V12, and raises MAP from 0.574 to 0.618 on V13, confirming clear gains in top- n recommendation quality.
Ali et al. (Mon,) studied this question.