A machine learning-based approach for predicting a recipe’s cooking time from its ingredients and direction

Abstract

This paper focuses on predicting cooking times for recipes based on advanced natural language processing techniques and machine learning models. With the increase in the number of online recipe datasets, there is a massive opportunity to predict cooking times based on ingredients and cooking instructions, which will help the user make the right choices during meal preparation. The study uses recipe-specific embeddings derived from various NLP models, such as BERT-base, RecipeBERT[6], and CookBERT[8], to extract meaningful features from recipe texts, including ingredients and directions. These embeddings are used with regression models to predict cooking time. The dataset that is used in this study is 25,928 recipes extracted from allrecipes.com, and various preprocessing steps have been undertaken, such as text cleaning, time conversion, and outlier removal. Various machine learning algorithms, including Linear Regression, Random Forest, XGBoost, and Multi-Layer Perceptrons (MLP), have been experimented on this task of predicting cooking times. The experiment results prove that domain-specific embeddings, such as RecipeBERT and CookBERT, surpass general BERT models; in fact, the best model reaches an R2 score of 82.12% and a mean absolute error of 11.48 minutes. This study shows that domain- specific models can improve the ability of the semantic meaning captured between the cooking instructions and the ingredients to predict cooking time better. Moreover, applying the PCA, SVD, and UMAP dimensional reduction techniques that actually improved model performance by providing better data learning for this system has practical application in meal planning and selection of recipes by helping a user manage cooking time effectively.