Recipe recommendations for individual users and groups in a cooking assistance app

Abstract

Recommender systems are commonly-used tools to assist people in making decisions. However, most research has focused on the domain of recommendations for audio-visual content and e-commerce, whereas the specific characteristics of recommendations for recipes and cooking did not receive enough attention. Since meals are often consumed in group (with friends or family), there is a need for group recommendations, taking into account the preferences of all group members. Also cuisine, allergies, disliked ingredients, diets, dish type, and required time to prepare are important factors for recipe selection. For 13 algorithms, we evaluated the recommendations for individuals and for groups using a dataset of recipe ratings. The best algorithm and a baseline algorithm based on popularity were selected for our mobile kitchen experience and recipe application, which assists users in the cooking process and provides recipe recommendations. Although significant differences between both algorithms were witnessed in the offline evaluation with the dataset, the differences were less noticeable in the online evaluation with real users. Because of the cold-start problem, the advanced algorithm failed to reach its full accuracy potential, but excelled in other quality features such as diversity, perceived usefulness, and confidence. We also witnessed a better evaluation (about half a star) of the recommendations by the more advanced cooks.