Promoting Healthy Food Choices Online: A Case for Multi-List Recommender Systems

Abstract

Many food choices are made online. Interactive, personalized interfaces, such as recommender systems, can help users to ifnd new products to eat or recipes to cook, but they tend to promote unhealthy alternatives. In this position paper, we argue that better algorithms are not the only way forward. We blend algorithms and user interface design to present a multi-list recommender interface that presents multiple lists of personalized items in a single interface, where each list is optimized for a specific feature (e.g., 'less fat'). We argue how multi-list recommenders can be used to support healthier food choices.

Keywords

• eol>Recommender Systems
• Food
• Digital Health
• Interfaces
• Algorithms

1. Introduction

content is recommended (cf. [ 9 ]). We propose a multi-list recommender interface to support healthy food choices, Recipe websites have become increasingly popular. This based on the UI of Netflix [ 10 ]. This interface combines has spurred the development of food recommender sys- diferent ‘single lists’ of algorithms that use a specific optems, which help users to navigate the thousands of timization (e.g., ‘Drama movies’ or ‘Because you watched recipes found online, by presenting personalized content Frozen’) to come up with a new subset of similar recom[ 1 ]. However, online users struggle to identify healthy mendations, which are combined into a comprehensive recipes, due to a lack of knowledge and misleading cues, multi-list UI. We describe how this can be applied in the and because popular recipes tend to be unhealthy [ 1 ]. food domain to promote healthier alternatives, taking as

An increasing number of algorithmic approaches to a starting point an interface that suggests alternatives personalization have surfaced [ 2 ]. For example, recom- for a recipe that a user had searched for, either through mender systems present food (e.g., meal plans) that is a look-up task or implicitly in an exploratory search. similar to what a user liked in the past [ 1 ]. Yet, there has been little attention for health [ 1, 3 ]. It has been shown that computing similarity between recipes (cf. [ 2 ]) is 2. Case: Multi-list Recommenders not enough to shift preferences towards health-related behaviors [ 3, 4 ]. In fact, it is suggested that suitable rec- Most recommender systems optimize their content to be ommendations for one’s current preferences could even similar towards a user’s past preferences. Algorithmibe counterproductive [ 5, 6, 7 ], especially if one’s current cally speaking, this leads to “more like this” recommenlifestyle is rather unhealthy [ 3 ]. dations [ 1 ]. For example, in the food domain, if a user has

Most users stick to familiar recipes [ 8 ]. For example, in bookmarked several recipes that contain potatoes, then the context of sustainability, users might swap their ham- a recommender system will present more potato-based burger’s beef-based patty for a plant-based one, relying recipes. The downside of this approach is that if a user on familiar substitutes. Although this is consistent with is currently unhealthy, she is reinforced into her current recommendation strategies on recipe websites, which preferences through more unhealthy content [ 3, 4 ]. typically present a ‘more like this’ set of recommenda- A few solutions have been proposed to alleviate this. tions alongside each recipe [ 2 ], such similar recommenda- For example, imposing health constraints on the algotions are unlikely to produce healthy recommendations. rithm might boost the healthiness of chosen items [ 11 ],

In this position paper, we argue that personalization but this might leave users dissatisfied if they have no approaches should go beyond only changing what is rec- healthy eating goals. Moreover, one could show a much ommended, by also focusing on the decision context: how larger list of recommendations, but this will lead to a sharp increase in choice dificulty, or choice overload, Joint Proceedings of the ACM IUI 2021 Workshops, April 13–17, 2021, when not explained properly to a user [ 12 ]. College Station, USA We argue that a multi-list recommender system can ("C. aTlaraintt.sntearr)[e-mail] (A. D. Starke); [e-mail] overcome algorithmic biases towards unhealthy foods © 2021 for this paper by its authors. Use permitted under Creative Commons and mitigate choice overload. It comprises novel algoCPWrEooUrckReshdoinpgs IhStpN:/c1e6u1r3-w-0s.o7r3g LCicEenUseRAttWribuotironk4s.0hIontpernPatrioonacl e(CeCdBiYn4g.0)s. (CEUR-WS.org) rithmic and interface components. In terms of content, a multi-list recommender uses multiple algorithms that each retrieve similar content, but difer in terms of what specific attributes are optimized for. For example, similarity could be computed across multiple attributes (e.g., ingredients, nutrients) to optimize utility (cf. [ 13 ]), of which the importance weights can be adapted in line with a user’s goals (e.g., putting more weight on calorie content for users who wish to lose weight). In terms of the interface, the most important contribution is to highlight the diferences between the diferent ‘single lists’, by explaining them to the user. For example, by highlighting the similarity of new recommendations with items liked in the past [ 10 ] or by emphasizing health benefits in a food recommender system [ 6 ].

A recommender study on computers shows how this can be achieved, by using a ‘critiquing’ approach [ 14 ].

The recommender system would present an initial item that fits a user’s past preferences, for which alternatives are presented underneath it across various attribute categories, such as ‘Cheaper and Heavier’ or ‘Higher Processor Speed, but More Expensive’. This way, each additional attribute category forms a new list of similar recommendations, which are diversified based on diferent attributes, as well as explained clearly to a user.

An example of a multi-list recommender interface is depicted in Figure 1. A reference recipe is depicted at the top, which a user may have searched for earlier. To explore alternatives, similar item recommendations are presented underneath it, in separate ‘single lists’, where each list is optimized towards a certain attribute. In our example, the first row optimizes for similarity with the reference recipe, explained as “Recipes that contain similar ingredients”. This could be achieved by using similarity functions and metrics, such as cosine similarity and RMSE [ 2 ]. In contrast, the second row focuses on ‘Similar recipes with less fat’, which could be achieved by either putting more attribute weight on a recipe’s fat content or determining the top-100 similar recipes and subsequently re-ranking that list on fat content. This way, multi-list recommender interfaces can be used to support a variety of user goals, recommending recipes that people like, yet supporting variations that some users might find particularly interesting. 2.1. Directions for Future Research The efectiveness of multi-list recommenders to support changes in user preferences and behavior has yet to be tested. The food domain is an excellent starting point, for multi-list interfaces provide algorithmic diversity that is needed to improve current unhealthy approaches in food recommenders [ 3, 4 ]. We propose two lines of research to examine visual UI design in recommender systems.

First, diferent list representations (i.e., single list vs multilists) should be compared in terms of choice behavior and user evaluation, for it is currently unclear to what extent they are efective [ 10 ]. For example, does a multilist representation reduce choice overload? Second, we propose to investigate what types of explanatory labels in a multi-list representation are the most persuasive to shift a user’s food preferences. For example, should they highlight health prevention (‘Similar, but less fat’), or health promotion attributes (‘Similar, but more fiber’)?

Another interesting avenue of research is the use of visual cues, which can afect consumer preferences for food selection [ 15, 16 ]. For one, product packaging color may be adapted to invoke certain emotions in supermarkets [ 16 ]. In an online context, recipe websites could exploit ifndings that recipes tend to be rated more favorably if they are accompanied by visually attractive photos [ 15 ].

Above all, we argue to evaluate multi-list recommender systems through a user-centric approach (cf. [ 17 ]). Not only should be examined what recipe is chosen, but also how users perceive and evaluate a multi-list recommender interface, for instance, compared to a single list approach. An important measure would be the perceived choice dificulty (cf. [ 12 ]), since it is currently unclear whether presenting many sub-lists (5+) is possible without afecting user satisfaction. Moreover, it would be interesting to examine which types of ‘single lists’ are the most efective in supporting health food choices, which could be a interdisciplinary field of study between computer science and nutrition science. In doing so, it would be important to striking the right balance between algorithmic accuracy (i.e., reducing RMSE) and interface design (e.g., nudges).

3. Conclusion

We have proposed how healthy food choices could be supported by combining adaptations in the presented content (i.e., algorithms) and the decision context (i.e., the interface), in a new approach. We have argued how to enable users to find and to select healthy content in a recommender system. Fundamental to our approach is that users are still given the freedom to choose what they want, in line with research on nudging [ 9 ], but that the use of the interface would not trigger choice overload or increase choice dificulty [ 12 ].

For future research, we stress that interfaces and algorithms are not two mutually exclusive categories of research. Our multi-list recommender systems case nicely illustrates how ‘similar content’ and ‘healthy content’ can go hand in hand, by pointing out what each single list of recommendations represents. We expect that diversifying the diferent types of recommendations presented, rather than only focusing on algorithmic optimization will be more efective in supporting healthy eating habits. For instance, should these lists always be fully personalized, or can they be less personalized in terms of past preferences and optimized to a user’s eating goals?

4. Acknowledgments

This work is in part funded by MediaFutures partners, the Research Council of Norway (grant number 309339), and the Niels Stensen Fellowship.

References

[1] C. Trattner , D. Elsweiler , Food recommender systems: Important contributions, challenges and future research directions , arXiv preprint arXiv:1711.02760 ( 2017 ).

[2] C. Trattner , D. Jannach , Learning to recommend similar items from human judgments, User Modeling and User-Adapted Interaction ( 2019 ) 1 - 49 .

[3] A. Starke , Recsys challenges in achieving sustainable eating habits , in: HealthRecSys'19: Proceedings of the 4th Workshop on Health Recommender Systems, ACM , 2019 , pp. 29 - 30 .

[4] C. Musto , C. Trattner , A. Starke , G. Semeraro, Towards a knowledge-aware food recommender system exploiting holistic user models , in: Proceedings of the 28th ACM Conference on User Modeling, Adaptation and Personalization , 2020 , pp. 333 - 337 .

[5] M. D. Ekstrand , M. C. Willemsen , Behaviorism is not enough: better recommendations through listening to users , in: Proceedings of the 10th ACM Conference on Recommender Systems , ACM, 2016 , pp. 221 - 224 .

[6] C. Musto , C. Trattner , A. Starke , G. Semeraro, Exploring the efects of natural language justifications on food recommender systems , in: Proceedings of the 29th ACM Conference on User Modeling, Adaptation and Personalization , 2021 .

[7] A. D. Starke , M. C. Willemsen , C. Snijders , With a little help from my peers: Depicting social norms in a recommender interface to promote energy conservation , in: Proceedings of the 25th International Conference on Intelligent User Interfaces , IUI '20 , 2020 , p. 568 - 578 .

[8] Y. M. Asano , G. Biermann, Rising adoption and retention of meat-free diets in online recipe data , Nature Sustainability 2 ( 2019 ) 621 - 627 .

[9] R. H. Thaler , C. R. Sunstein , Nudge: Improving decisions about health, wealth , and happiness, 2009 .

[10] C. A. Gomez-Uribe , N. Hunt , The netflix recommender system: Algorithms, business value, and innovation , ACM Transactions on Management Information Systems (TMIS) 6 ( 2015 ) 1 - 19 .

[11] C. Trattner , D. Elsweiler , Investigating the healthiness of internet-sourced recipes: implications for meal planning and recommender systems , in: Proc. of WWW '17 , 2017 , pp. 489 - 498 .

[12] D. Bollen , B. P. Knijnenburg , M. C. Willemsen , M. Graus , Understanding choice overload in recommender systems , in: Proceedings of the fourth ACM conference on Recommender systems, ACM , 2010 , pp. 63 - 70 .

[13] B. P. Knijnenburg , M. C. Willemsen, Understanding the efect of adaptive preference elicitation methods on user satisfaction of a recommender system , in: Proceedings of the third ACM conference on Recommender systems , 2009 , pp. 381 - 384 .

[14] P. Pu , L. Chen, Trust-inspiring explanation interfaces for recommender systems , Knowledge-Based Systems 20 ( 2007 ) 542 - 556 .

[15] A. Starke , M. Willemsen , C. Trattner , Nudging healthy choices in food search through visual attractiveness , Frontiers in Artificial Intelligence. Preprint ( 2020 ).

[16] I. Vermeir , G. Roose, Visual design cues impacting food choice: A review and future research agenda , Foods 9 ( 2020 ) 1495 .

[17] B. P. Knijnenburg , M. C. Willemsen , Evaluating recommender systems with user experiments , in: Recommender Systems Handbook , Springer, 2015 , pp. 309 - 352 .