<![CDATA[Hotel recommendation systems often fail to recommend new hotels due to extreme data sparsity problems (item cold-start) and are vulnerable to the computational over-smoothing phenomenon. This study aims to comprehensively evaluate and compare the ranking architectures of Graph Convolutional Network (GCN) and LightGCN. The method used is a computational experiment using an Ablation Study approach to dissect the effect of propagation depth (1-Layer vs. Multi-Layer) and the injection of external features. The evaluation was conducted on a small-sized dataset from Mendeley Data containing review histories for 16 hotel entities. The main results show that the 1-Layer LightGCN with features is the most superior model for active users (warm-start), achieving an NDCG@10 score of 0.8702. However, in the extreme new hotel scenario (0-shot cold-start), this shallow architecture failed, and the best solution was actually won by the pure Multi-Layer model without features (No-Feature), which achieved a Hit Ratio (HR@10) of 71.43%. In conclusion, there is no single perfect model for all conditions; system implementation is recommended to adopt a dual-framework that integrates the speed of 1-Layer LightGCN and the propagation robustness of the Multi-Layer model.]]>
Copyrights © 2026
