Enabling Better Healthy Food Recommendation By Incorporating Contextual Knowledge

Food recommender systems (FRS) are increasingly recognised as valuable tools for simplifying food decision-making, while also promoting healthier eating habits. However, accurately predicting user preferences in this context remains a challenging task: choosing what to eat is a multifaceted and highly complex process. The greater challenge lies in recommending food that users will not only enjoy, but that is also healthy, requiring work that goes beyond the typical recommender system (RS) algorithms, which focus on maximising expected ratings and typically may encourage people to make less healthy choices.

This thesis seeks to improve food recommendation performance and address the taste-healthiness trade-off challenge and by integrating several forms of contextual knowledge, including situational context. Very little extant research has systematically explored the impact of multiple dynamic factors on influencing people's eating habits, recipe ratings, and nutritional intake behaviours. Most existing research focuses on single contextual factors, typically simple extrinsic ones such as location and time.

This research addresses these gaps by understanding daily eating habits through semi-structured interviews, followed by a large-scale experimental study in simulated contexts to examine how these contexts influence recipe rating behaviour and, therefore, implied nutritional intake. The results highlight the importance of developing context-aware recommender systems (CARS) leading to the development of novel one-stage and two-stage contextual modelling approaches with multimodal feature sets. An innovative method for generating healthy recommendations and a novel evaluation approach are also proposed.

Key findings suggest that dynamic contextual factors like emotions, busyness, seasons, and physical activities influence food choices and decision-making. People's preferences for healthy recipes vary with context, especially under stress. Incorporating such contextual features into RS algorithms significantly improves recipe rating predictions and permits healthier recommendations to be made. A contextual healthy recommendation approach and a novel evaluation metric, RMSEh, are introduced to align recommendations with individual preferences whilst promoting healthier choices.

This research has implications for the future development of FRS, and shows that emotion-aware systems could lead to better healthy food recommendations. These findings provide valuable insights into the design of more sophisticated healthy food recommender systems and offer a promising framework for the development of context-aware recommender systems across various application domains.