Health-Aware Travel Recommendation Through Environmental Risk Profiling and Carbon-Aware Model Selection

Abstract

The carbon footprint of frequent model retraining is a growing component of recommendation system operating cost, and travel recommendation systems rarely incorporate destination-level health information for users with chronic conditions. This paper presents a methodological framework that integrates both concerns at the feature engineering and model-selection stages of a tabular recommender pipeline. The construction draws on 980 real user preference profiles from the UCI Travel Review Ratings dataset, 81 globally distributed destinations enriched with elevation, climate, ultraviolet, and air-quality measurements retrieved at runtime from Open-Meteo, and condition-specific environmental thresholds grounded in clinical guidance for asthma and chronic obstructive pulmonary disease, cardiovascular disease, pregnancy, photosensitivity, and diabetes. Health conditions are sampled at WHO-aligned prevalence rates, and the recommendation label is constructed as a transparent weighted combination of preference match and health safety with additive Gaussian noise; the design discussion in Section 3.4 motivates this choice in the absence of a public dataset linking real user travel preferences with real health-condition records. Eleven machine learning models are evaluated across predictive accuracy, ranking quality, training time, energy consumption, and CO₂ emissions over three random seeds. XGBoost reaches the lowest RMSE of 0.266 and the highest NDCG@10 of 0.983 while emitting 0.007 g CO₂ per training run. The 200-tree Random Forest reaches a slightly worse RMSE of 0.287 while emitting 0.152 g, an emission ratio of approximately 22 with no measurable accuracy improvement. On the present benchmark, gradient boosting therefore Pareto-dominates Random Forest at every complexity level evaluated, a pattern consistent with the wider tabular-data literature. The health-aware feature pipeline shifts top-ten recommendations for non-healthy subpopulations away from destinations whose environmental profile exceeds condition-specific clinical thresholds, providing an implicit ablation against a preference-only baseline. The framework is released as an open-source replication package, and the present results should be read as a structured engineering proof of concept rather than a clinical validation.