Leakage-aware hybrid tabular deep learning with target-provenance auditing for calorie-expenditure prediction and constrained exercise food recommendation

Abstract

Abstract Personalised exercise planning depends on calorie expenditure estimates derived from valid predictors. Published tabular recommenders rarely scrutinise the provenance of their outcome variable. We propose a leakage-aware, audit-first framework that combines hybrid tabular deep learning with a constrained exercise-food recommender, and apply it to three public tables with 3,864 workout records, 400 catalogued foods, and 512 user-food ratings. A pre-modeling audit found and removed direct target leakage that had inflated a previously reported R 2 to 0.998. Formal distributional forensics determined that the outcome column is statistically indistinguishable from Uniform (100, 500) noise (Kolmogorov–Smirnov p = 0.489), which corresponds to an analytic mean-absolute-error floor of 100.00 kcal. We trained three corrected architectures, a Tabular 1D-CNN, a feature-token Transformer and a BiLSTM plus feature-token Transformer, under record-level 64/16/20 partitioning with training-only augmentation against six conventional learners and a median baseline. The best model produced a mean absolute error of 101.68 kcal on 773 unseen test instances, within 1.7% of the theoretical lower bound, and repeated cross-validation, bootstrap confidence intervals, paired Wilcoxon tests, subgroup auditing across 29 strata, and seven families of explainability analyses all agreed on the same conclusion. The deterministic recommender achieved 80.0% profile coverage, 100% catalogue coverage and no constraint violations over 240 simulated profiles and 1200 emitted candidates, with all residual target error attributed to the predictor rather than the ranking logic. The framework provides a transferable protocol for validating outcome variables prior to modelling, and a reproducible, auditable recommendation pipeline for exercise and nutrition informatics.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗