Spatial Graph-Enhanced Accommodation Demand Forecasting Using KD-Tree: Algorithm-Agnostic Validation with Explainable AI

Abstract

Accurate accommodation demand forecasting is critical for revenue optimization and resource allocation in the hospitality industry. This study proposes a spatial graph-based forecasting framework that constructs a KD-Tree proximity network between 1,455 hotel properties and 452,044 tourism activities sourced from the Travlr Challenge dataset. PageRank and Degree Centrality are extracted from the resulting bipartite graph comprising 8,979 nodes and 43,650 edges as supplementary predictive features for ensemble models. Validation employs a 2x3 algorithm-agnostic experimental matrix encompassing LightGBM, XGBoost, and Random Forest under Baseline and Proposed architectures, with 5-Fold Cross-Validation enforcing zero label leakage. An ablation study further isolates the contribution of graph-derived topological features against four simpler spatial proximity baselines to confirm genuine structural predictive signal. LightGBM Proposed achieves the lowest RMSE of 1.1684 with a statistically significant improvement over Baseline confirmed by Paired T-Test with p-value of 0.0103, representing a 3.23% reduction. SHAP Out-of-Fold analysis identifies PageRank as the dominant predictor with a mean SHAP value of 0.2273, surpassing all intrinsic and spatial proximity features, while Degree Centrality contributes zero predictive signal once PageRank is available. The ablation study confirms that predictive gains originate from graph-derived structural information rather than general location effects. Graph feature extraction imposes negligible computational overhead of less than 0.15% of total pipeline time, confirming production-scale viability and direct integrability into existing revenue management systems.