Source-linked AI summary
A Machine Learning Framework for Predicting Restaurant Food Waste to Support Sustainable Food Management
Md Mehedi Hasan Naeem, Md Ashraful Islam, Moumita Barua, Ishtiyak Ahmmad Araf, Md. Arefin Haque Mahir
TL;DR
Restaurant food waste presents major sustainability and economic challenges, while large-scale measured waste data are unavailable. This paper develops and evaluates a reproducible machine-learning framework using a constructed target; ensemble methods performed best, with Random Forest reaching MAE= 6.19 kg, RMSE= 8.36 kg and R2 = 0.817 against that proxy target.
Problem
Restaurant food waste creates substantial environmental, economic and social consequences, but large-scale ground-truth food waste measurements are not publicly available.
Method
The paper constructs a reproducible 77,980-record dataset and proxy target, then compares four regression models using chronological train-test evaluation and 5-fold time-series cross-validation.
Results
Ensemble methods consistently outperformed linear baselines; Random Forest achieved MAE= 6.19 kg, RMSE= 8.36 kg and R2 = 0.817 against the constructed proxy target.
Takeaways & Limitations
The realistic pre-service feature subset contains meaningful predictive signal for the constructed target, while the disclosed dataset, formula and code support reproducibility and future extension to measured waste data.
Takeaways & Limitations
All reported metrics measure fit to an author-constructed proxy target rather than validation against empirical food waste measurements.
Abstract
from arXiv · showhide
Food waste in the restaurant sector poses a substantial challenge to environmental sustainability and economic efficiency. This paper presents an exploratory machine learning framework for estimating daily restaurant food waste quantities from operational and contextual features. A structured dataset was constructed by integrating restaurant demand records, meteorological data and temporal event indicators, yielding 77,980 records across 27 features. Because large-scale ground-truth food waste measurements are not publicly available, the target variable was derived from operationally justified assumptions, with the complete construction formula and controlled stochastic variability disclosed for full reproducibility. Four supervised regression models, namely Linear Regression, Decision Tree, Random Forest and Gradient Boosting, were evaluated under a chronological 70-30 train-test split that respects the temporal ordering of restaurant operations, augmented by 5-fold time-series cross-validation. All reported metrics are explicitly scoped to performance against the constructed target and do not imply validation against measured food waste. Ensemble methods consistently outperformed linear baselines. Random Forest attained an MAE of 6.19 kg, RMSE of 8.36 kg and $R^2$ of 0.817 on the realistic feature subset following systematic exclusion of algebraically leakage-prone variables. Feature importance analysis identified menu diversity, operational area and temporal activity patterns as the primary predictive drivers. The full dataset, target construction formula, codebase and experimental configurations are publicly released to support reproducibility and future extension to empirically measured waste data.
I. INTRODUCTION
The paper introduces a reproducible machine-learning framework for estimating restaurant food waste from operational and contextual data. It combines a public dataset, leakage-aware evaluation, model comparison, and analysis of predictive drivers.
- Contributions: Four supervised regression models are evaluated using a chronological train-test split and 5-fold time-series cross-validation.The evaluation preserves the temporal ordering of restaurant operations.
- Contributions: Feature leakage analysis identifies algebraically entangled variables as a source of inflated performance estimates.The study explicitly examines leakage sources and quantifies their impact through ablation.
- Contributions: Menu diversity, operational context, and temporal activity are identified as important drivers of the predicted waste target.The framework focuses on operational and contextual predictors relevant to restaurant sustainability.
- Contributions: The open-source framework is intended to support future research using empirically measured food-waste data.The paper positions reproducibility and future extension as explicit contributions.
II. RELATED WORK
The paper situates restaurant food-waste prediction within broader food-service forecasting research and describes a dataset-building approach based on integrated operational, meteorological, and temporal data. Because measured waste data are unavailable at scale, the target is constructed from an operational formula rather than direct observations.
- Research context: Prior work addresses food demand and sales forecasting, but does not directly estimate restaurant food waste.The cited examples include seasonal and quantile-based forecasting approaches for food sales and supermarket demand.
- Dataset construction: 77,980 daily records across 27 features combine restaurant demand records, meteorological measurements, and temporal event indicators.The sources are aligned by date and processed through imputation, duplicate removal, categorical encoding, and scaling.
- Target construction: The target food_waste_kg is estimated with an operationally grounded formula because no large-scale publicly available dataset of directly measured restaurant waste exists.The formula uses estimated preparation, consumption ratio, calibration, efficiency, and Gaussian-noise terms.
- Target construction: Metrics evaluate fit to the constructed target and should not be interpreted as validated predictions of real-world food waste.The framework is explicitly presented as an exploratory methodology baseline for future measured-data research.
C. Feature Summary
The feature-summary section describes the retained dataset features and the broader preprocessing workflow. It covers feature retention, data integration, preprocessing, and temporal-feature extraction.
- Feature retention: The retained feature set is summarized after leakage reduction for model development.This identifies the feature summary as post-selection rather than a description of the original full feature set.
- Workflow: The framework pipeline includes data integration, preprocessing, exploratory analysis, feature selection, model training, and evaluation.Figure 1 is cited as an overview of this workflow.
- Preprocessing: Categorical variables are ordinally encoded, while continuous operational and pricing features are standardized to zero mean and unit variance.Missing values and duplicate records are removed before these transformations.
- Temporal features: Temporal features extracted from dates include year, month, ISO week number, and day-of-week indicators.These features are used to represent seasonal and cyclical patterns associated with restaurant waste generation.
B. Exploratory Data Analysis
Exploratory analysis identified menu diversity and operational area as the strongest positive associations with the constructed waste target, while temporal and contextual variables provided additional signal. The study also addressed structural leakage by excluding variables algebraically entangled with target construction and compared four regression models.
- Menu diversity and operational area exhibit the strongest positive associations with the target, while temporal and contextual indicators provide moderate but consistent signal.
- Feature leakage: Full-feature experiments produced R2 values exceeding 0.99 because food_prepared_kg and food_sold_kg participate directly in target construction, while num_orders is an algebraic antecedent.
- Feature leakage: food_prepared_kg, food_sold_kg and num_orders were excluded from the realistic feature subset to reduce structural and indirect leakage.
- Regression models: Four supervised regression models were evaluated: Linear Regression, Decision Tree, Random Forest and Gradient Boosting.
E. Chronological Train-Test Split and Cross-Validation
The evaluation preserves the temporal ordering of restaurant operations by training on earlier records and testing on later records. An expanding-window 5-fold procedure additionally assesses stability across successive chronological training windows.
- The dataset was split at the 70th percentile of the date range, using the earlier 70% for training and the remaining 30% for testing.
- The chronological split avoids allowing future observations to inform predictions of past records and provides a more realistic prospective forecasting evaluation.
- 5-fold TimeSeriesSplit cross-validation uses expanding chronological training sets and evaluates each model on the immediately following period.
F. Evaluation Metrics
Performance was assessed with complementary regression metrics that quantify average error, sensitivity to large errors and explained target variance. Because the target is constructed, R2 measures fit to the defined proxy rather than validated real-world food waste prediction.
- MAE reports the average magnitude of prediction errors in kilograms, providing an interpretable measure of practical accuracy.
- RMSE penalizes large individual errors more heavily than MAE and is therefore sensitive to outlying predictions.
- R2 measures the proportion of target variance explained by the model.
- R2 should be interpreted as model fit to the constructed proxy rather than evidence of real-world food waste prediction capability.
V. EXPERIMENTAL RESULTS AND DISCUSSION
On the realistic feature subset, ensemble methods outperformed both baselines in chronological held-out testing, with Random Forest achieving the strongest reported performance against the constructed proxy target.
- The evaluation used a chronological 70-30 split on the realistic feature subset, with metrics evaluated against the constructed proxy target.
- Ensemble methods outperformed both baselines across all metrics on the realistic feature subset.
- Random Forest achieved the best result, reducing MAE by 30.3% relative to Linear Regression.
- Figure 4 compares R2 scores for the evaluated models on the held-out chronological test set.
B. Cross-Validation Results
Five-fold time-series cross-validation assessed performance across successive chronological training windows and found Random Forest had the highest mean R2, while Gradient Boosting was the most stable nonlinear model.
- Random Forest achieved the highest mean cross-validation R2 of 0.720.
- Gradient Boosting showed the lowest variability among nonlinear models, with a standard deviation of 0.032.
- The held-out Random Forest test performance exceeded its cross-validation mean, indicating variation across successive temporal windows.
- Feature Importance: Feature importance used impurity-based mean decrease in node impurity accumulated across all Random Forest trees.
- Feature Importance: Menu diversity and operational footprint were the two most influential features, while weekend, holiday, pricing, and promotional variables contributed less.
D. Ablation and Leakage Analysis
Ablation analysis identified structural and indirect leakage in target-related variables; excluding num_orders reduced Random Forest R2 to 0.817 on variables available before service commencement.
- The stepwise ablation provided a transparent account of how target-related variables influenced reported performance.
- Removing food_prepared_kg and food_sold_kg reduced R2 by 0.007 because they participate directly in target construction.
- num_orders remained an indirect leakage source because it encoded the order volume used to derive food_prepared_kg.
- Removing num_orders reduced R2 to 0.817 on the reduced feature subset available before service commencement.
- Figure 6 visualizes constructed target values against Random Forest predictions on the chronological held-out test set.
E. Discussion
The realistic feature subset contains predictive signal for the constructed target, while the study’s conclusions remain bounded by proxy-target evaluation and limited generalizability evidence.
- The realistic feature subset contained meaningful predictive signal for the constructed food waste target.
- Ensemble methods captured nonlinear interactions among operational, contextual, and temporal variables.
- All reported metrics fit an author-constructed proxy target rather than validating predictions against empirical food waste measurements.
- Walk-forward validation using unseen restaurant locations would provide a stronger test of temporal and geographic generalizability.
- Deployment would require integration with point-of-sale, kitchen-management, and direct waste-monitoring systems to replace the proxy-derived target.
- The study reports 77,980 records, publicly released resources, and Random Forest performance of MAE= 6.19 kg, RMSE= 8.36 kg and R2 = 0.817 against the proxy target.