The exponential growth of scholarly publications has created significant challenges for researchers, requiring recommender systems that address multiple competing objectives beyond accuracy alone. This paper provides a systematic comparison of algorithmic methods of scholarly recommendation based on the evaluation of 38 studies from 2017 to 2024. Three key dimensions of novelty, serendipity, and integration of multilayer knowledge graphs are quantitatively defined to set up these comparative standards. The results discover RWR to be optimal to recommend awareness of novelty, which improved exploration measures by 15–25%; Bisociative Information Networks (BisoNets) emerge as optimal for facilitating serendipitous discovery, achieving a 25–35% cross-disciplinary recommendation rate; and the multilayer knowledge graph as optimal to envisage complete recommendations as it improved accuracy by 20–30%. Based on these findings, an evidence-based selection framework is proposed that maps system priorities to optimal algorithm selection, underpinned by a rationale of justification and computational complexities. Such a framework even provides workable instructions for hybrid implementations that address the inherent multi-objective optimisation challenge of academic discovery, as well as offers implementable lessons for designing next-generation scholarly recommenders.