<![CDATA[Estimating portion sizes and cooking times are goals for smart kitchen assistants, enabling better meal planning and reducing food waste due to over-portioning. Existing approaches in computational gastronomy often struggle to provide estimates from prepared ingredient data. This study uses XGBoost to extract features from a dataset containing 1,400 Indonesian recipes to predict binary classification targets for portion sizes and required cooking times. The dataset used for the prediction includes information on ingredients and their quantities, as well as preparation steps. In addition to the recipe dataset, the TKPI dataset is also used to help determine the category of food ingredients, protein content, and cooking technique complexity. This dataset is then further optimized with hyperparameters to maximize model performance. This paper conducted trials with 6 models where the best model for portion size had an accuracy of 0.7821 with a balanced accuracy of 0.4929, and an F1 Score of 0.8763, while the accuracy for cooking time was 0.6929 with a balanced accuracy of 0.6445, and an F1 Score of 0.7737. From the best model, it was found that the quantity of weighted ingredients and the distribution of ingredients per step were among the most influential features, while step-based and technique-based features were the most important features for cooking time. The contribution of this research is the development of an interpretable model for meal planning efficiency in culinary applications. These results indicate that feature aggregation combined with XGBoost provides actionable insights for smart kitchen assistants and recommendation systems.]]>
