A Graph-Based Article Recommendation System with Multilayer Citation
Makaleler > Tam Makale · 2026
Yazarlar: Umut Karadurmus, Züleyha AKUSTA DAĞDEVİREN
International Journal of Multidisciplinary Studies and Innovative Technologies, Uluslararası, 2026, cilt 10, ss 64-64
Özet
We present a citation-centered article recommendation system that ranks works in a heterogeneous two-hop OpenAlex graph by combining structural connectivity with concept overlap and venue signals. Direct references form a primary, local-reading tier, whereas references reached through them form a secondary discovery tier. Ten target articles were evaluated from a data snapshot collected on March 8, 2026. Because concept-set Jaccard is also a scoring input, we supplemented it with an independent abstract-level TF–IDF cosine analysis and a leave-one-out hidden-reference experiment. For the full model, mean abstract cosine was 0.045 for primary and 0.029 for secondary recommendations among available abstract pairs. Across 358 held-out direct citations, 45.5% re-entered the two-hop candidate pool; Recall@10 was 0.112 overall and 0.245 when conditioned on candidate coverage. The networks were sparse and weakly connected when citation direction was ignored, although these properties are partly induced by ego-network construction. The findings show that a lightweight and interpretable hybrid can recover relevant citation-neighborhood works without model training, while also exposing candidate-generation and ground-truth limitations that require larger labelled or user-based evaluation.