Author: Prof. Xu Wang
Email: wangxu02@nuit.edu.cn
Institution: Neusoft Institute Guangdong School of Computer Science
We examine the operating boundary of preference-aware augmentation for a shared community ranker. A new, fully specified degree-corrected block simulation crosses alignment, missing social edges, homophily controls and proposal budgets, with 30 independent seeds per cell. Preference fusion can improve on a matched structural blend without an explicit preference-homophily multiplier, but this does not establish a unique benefit from two-hop eligibility. On Last.fm, GTC has NDCG@20 of 0.0545 versus 0.0511 for the observed graph and 0.0632 for unrestricted preference neighbors. On Delicious, the corresponding values are 0.0135, 0.0132 and 0.0099, while the matched structural blend reaches 0.0146. UserKNN and LightGCN outperform GTC on both observed datasets. Paired inference over all five Louvain seeds supports no uniform improvement over the observed graph. The revision also identifies sensitivity to simultaneous-event ordering, supplies a strict timestamp-block analysis, corrects zero-removal graph handling, and quantifies affinity corruption, preference drift and candidate-capping trade-offs. These results support

