Research Work: Calorie Estimation using Ensemble Technique from Recipes of Bangladeshi Foods
Research Approach: This research employs a text-based regression framework to estimate calorie values from Bangladeshi food recipes. A novel dataset of recipe texts and verified calorie information is used to train multiple transformer-based models. Their predictions are integrated using a stacking ensemble approach with Ridge regression as the meta-learner, enabling robust and accurate calorie estimation through effective semantic feature learning.
Result: The model evaluation demonstrates a promising performance across multiple standard metrics. With an MSE of 12,345.55 and an RMSE of 111.11, the model shows a strong ability to produce predictions close to actual values, while the MAE of 89.16 highlights its consistent reliability across individual data points. Although the R² value is near zero, this indicates a significant opportunity for further refinement and feature enhancement, emphasizing the model’s growth potential. The MAPE of 34.80% suggests that, on average, predictions remain within a practical and manageable range of actual outcomes, making the model suitable for real-world applications. Overall, these metrics underscore a solid baseline performance and provide a strong foundation for iterative improvement, indicating that with targeted optimizations, the model can achieve even higher predictive accuracy and practical utility.
Status: Accepted for publication at IEEE 2nd International Conference on Computing, Applications, and Systems (COMPAS 2025)
Research Paper: [Paper on IEEExplore]
Resources: GitHub Repository (Code and Dataset)