Innovative food recommendation systems: a machine learning approach

Abstract

Recommendation systems employ users history data records to predict their preference, and have been widely used in diverse fields including biology, e-commerce, and healthcare. Traditional recommendation techniques include content-based, collaborative-based and hybrid methods but not all real-world problems can be best addressed by these classical recommendation techniques. Food recommendation is one such challenging problem where there is an urgent need to use novel recommendation systems in assisting people to select healthy, balanced and personalized food plans. In this thesis, we make several advances in food recommendation systems using innovative machine learning methods. First, a novel recommendation approach is proposed by transforming an original recommendation problem into a many-objective optimisation one that contains several different objectives resulting in more balanced recommendations. Second, a unified approach to designing sequence-based personalised food recommendation systems is investigated to accommodate dynamic user behaviours. Third, a new food recommendation approach is developed with a temporal dependent graph neural network and data augmentation techniques leading to more accurate and robust recommendations. The experimental results show that these proposed approaches have not only provided a more balanced and accurate way of recommending food than the traditional methods but also led to promising areas for future research.