Hierarchical heterogeneous information networks and approximate reduction under semantic controllability

Real-world networks contain multiple entity and relation types, semantic paths, attribute nodes, and node features. Heterogeneous information networks (HINs) encode this heterogeneity, but placing all evidence in one graph space can obscure the distinct roles of relational structure and attributes and complicate controlled merge decisions. We propose the Hierarchical Heterogeneous Information Network (HHIN) as a data model that organizes typed entity relations in a main structure layer and descriptive evidence in a strong attribute layer. We develop controllable approximate reduction as one instantiation. Fixed point refinement grounded in behavioral equivalence yields stable structural candidates; two attribute Guards based on cosine similarity constrain merge admissibility; and a normalized semantic budget selects thresholds using similarity errors and normalized discounted cumulative gain at rank 10 (nDCG@10) computed from PathSim and HeteSim. Theory establishes termination, containment of accepted merges within stable candidate classes, and budget feasibility. On ACM, DBLP, and IMDB, the method reduces 7.22–21.78% of all nodes while retaining nDCG@10 values of at least 0.9735 for PathSim and 0.9910 for HeteSim. Across datasets, repeated retrieval and ranking yield speedups of 1.04–1.15 $$\times $$ and 1.13–1.26 $$\times $$ , respectively. On DBLP, paired tests with a relational graph convolutional network (R-GCN) and a heterogeneous graph attention network (HAN) do not detect significant classification differences between the original and reduced carriers, while mean epoch costs decrease by factors of 1.10 and 1.07. DBLP KMeans clustering and transductive label prediction from similarity neighborhoods provide additional downstream evaluations.