Recent developments in incremental learning for graph-based recommender systems from a complex network perspective

Incremental learning (IL) for graph-based recommender systems (GRS) must update models under memory and latency constraints, risking erosion of structural properties encoded in the interaction graph. This study examines IL for GRS through the lens of complex network theory, treating structural preservation – maintaining embeddings that approximate topological functionals such as communities, centrality and spectral quantities – as a primary objective. This study organises existing methods by the graph properties they implicitly preserve, review graph-centric evaluation metrics for diagnosing structural drift and outline open problems.