NutriHeal-Graph: Dataset for Goal-Conditioned Selective Verification of Nutrition-Goal Recipe Suitability

NutriHeal-Graph is a dish–goal pair dataset for nutrition-goal recipe suitability assessment under incomplete nutrition evidence. The dataset supports research in selective prediction, knowledge-driven dietary decision support, and graph-derived evidence integration for recipe recommendation. Each record represents a unique combination of a dish and one of four nutrition goals (G1–G4), labelled with a binary operational suitability indicator (target_strict_relevant) that equals 1 if and only if the dish satisfies all desired and avoided nutrient-indicator clauses of the goal. Labels are derived from KDRIs (Korean Dietary Reference Intakes) and WHO dietary guidelines, calibrated to a single-meal serving basis. Key statistics: 59,056 dish–goal pairs (14,764 unique dishes × 4 goals) 276 feature columns including quantity-aware APWIS graph-derived evidence features Train / Validation / Test split: 41,336 / 8,860 / 8,860 pairs (70/15/15) Overall positive rate: approximately 8.7% per split Source: Korean recipe platform (dish-level nutrient records) Nutrition goals: G1: High-protein with sodium/sugar control G2: Protein-supported sugar/carbohydrate control G3: Cardiovascular-oriented (high-protein, low-sodium, low-energy) G4: Weight-control balanced metabolic profile (high-protein, low-energy, low-carb, low-sugar) Nutrient indicator thresholds (single-meal serving): High energy ≥ 700 kcal, High protein ≥ 25 g, High carbohydrate ≥ 100 g, High sugar ≥ 15 g, High sodium ≥ 1,500 mg This dataset is intended for use with the accompanying experiment code (see linked code deposit). Leakage-prone columns (target_*, y_high_*, selection_*) are documented in excluded_columns.csv.