RecSys Challenges in achieving sustainable eating habits Alain Starke University of Bergen Bergen, Norway A.D.Starke@tue.nl ABSTRACT change can be achieved in the longer run [1, 3]. Most studies are limited to decision-making within an interface, while only few Most food recommender systems successfully fit one’s current studies monitor users over a longer time period [9]. However, appetite for different food items. However, one’s current straightforward, one-time web studies cannot tackle questions preferences (e.g. I like high-carb foods) may conflict with a new that concern changing habits, as this involves complex lifestyle dietary goal (e.g. I want to minimize my carb intake), which can aspects as social practices [2, 11]. Moreover, energy intervention be detrimental to recommendation quality over time. Moreover, studies in a persuasion or HCI context show that treatment choices in the short-term might not lead to behavioral change in effects can diminish quickly and may depend on the type of the longer-term. This short position paper outlines a few incentives provided [2], as superficial interface changes and challenges related to changing habits, and reports a small financial rewards only have short-lived effects. analysis to show how the Rasch model could complement CF- Another challenge is that one’s current habits and preferences based approaches in research on sustainable eating habits. may differ from one’s behavioral goals, such as when taking up a new diet. Traditional algorithms may simply reinforce one’s CCS CONCEPTS current behaviors (e.g. ‘perhaps you want a bag of crisps’), rather • Informa on Systems  Decision Support Systems than allowing users to discover new items that might have a • Human‐centered Compu ng  User Studies lower predictive accuracy, since they are at odds with one’s current habits (e.g. recommending low-calorie food). Although KEYWORDS tags may be able helpful for discovery of novel items [1], recommender systems might need to forgo on their predictive Behavioral Change; User Experience; Rasch Model; accuracy and focus on listening to users with regard to goal- Sustainability; Ea ng habits setting [3]. In addition, how decision psychology, persuasion and interface design can effectively complement such algorithms to 1 Introduction spur behavior change is currently an open question [12]. Recent data-driven studies have addressed the topic of sustainable eating [1, 13]. Whereas some users employ 3 Proposed research on changing habits recommender systems to acquire recipes that fit their appetite or Recent studies show key differences in the behavioral difficulty contain certain nutrients [9, 12], ‘sustainable eaters’ have a dual between various food items or energy-saving measures and their goal of fine dining at little environmental costs. To date, research resulting attractiveness [9, 10]. Whereas it is easy to achieve a has focused on algorithmic development (e.g. using CF-based small behavioral change (e.g. eating two cookies a day instead of RecSys), an approach that has proven its merit in domains where four), moving away from one’s current behavior is tricky. In a there is little difference between items with regard to the similar vein, behaviors related to medication adherence seem to behavioral thresholds to adopt or choose them, such as in movie vary in their execution difficulty [6], showing that fully sticking recommender systems. However, we argue that simply to one’s prescriptions is too hard for most patients. presenting items that resonate with one’s current habits and To take behavioral difficulty into account, Schäfer and preferences might not be effective in the long run [7]. Willemsen [9] and Starke et al. [10] have used the psychometric Rasch model to conceptualize nutrient intake and energy 2 Key challenges conservation as one-dimensional constructs. They show that A number of recommender domains show that presenting one’s motivation to perform a behavior (referred to as: ability or appropriate recommendations is only the first step of adoption attitude [5]) becomes apparent through the behavioral items one [3, 4, 13]. Although food, health, and energy recommender already engages in. This results in high adoption probabilities for systems present items that should ultimately result in behavioral items that are commonly performed (i.e. have a low behavioral change [3, 8, 10], there has been little work on how behavioral difficulty, ‘popular’), while such probabilities are low for obscure items (e.g. installing solar PV or optimizing fiber intake) [9, 10]. HealthRecSys’19, September 20, 2019, Copenhagen, Denmark. The interplay between a user’s ability and item difficulty © 2019 Copyright for the individual papers remains with the authors. Use permitted allows a recommender system to sketch the path towards under Creative Commons License Attribution 4.0 International (CC BY 4.0). behavioral change, particularly for the dual behavioral goal of 29 HealthRecSys’19, September 20, 2019, Copenhagen, Denmark A.D. Starke sustainable eating. The Rasch model suggests that it might be sustainable behavior, as both graphs show similar results. In more effective to take small steps over time, that slowly increase contrast, the Rasch-based set changes from mostly low-effort in behavioral difficulty and continuously match a user’s ability measures for low-ability users, to a mixed recommendation set [5, 11]. This would differ from CF-based approaches, as it does for high-ability users, including high kWh savings. not look users that have similar ‘bad’ habits, but relies on users that can lead the way to habit improvement. 4 Conclusion To exemplify the difference between a Rasch-based approach The Rasch model used in Figure 1 is one example of how an and CF-based approaches, we analyzed 7,551 dichotomous self- algorithm can consider changes in habits over predictive reports (‘yes’ or ‘no’) of 304 users to 134 energy-saving accuracy [cf. 7]. However, it does not address the question how a behaviors, including food measures. We predicted Top-10 user can effectively change his or her habits. The recommender recommendation sets for both approaches. Rasch would predict community should conduct more intervention studies over a the highest score for measures whose behavioral costs match a longer-term period, by not only focusing on optimizing the user’s ability. For the CF-based approach, we trained a Rating recommender’s predictive accuracy, but also by monitoring the Prediction model in MyMediaLite, using matrix factorization and user’s behavior. For example, while historical data could point 5-fold cross-validation, noting that the CF model was not very out one’s current habits, other methods of elicitation should accurate (RMSE = 0.49 under dichotomous data). reveal one’s behavioral goals. A useful method would be to conduct more studies using mobile devices, as it would allow Top-10 RecLists per cluster - Low Ability researchers to have user dialogs at multiple time points. 8 ACKNOWLEDGMENTS This work was made possible by funding from the Nielsen Mean Frequency 6 Stensen Fellowship. 4 REFERENCES [1] Asano, Y.M. and Biermann, G. 2019. Rising adoption and retention of meat-free diets in online recipe data. Nature Sustainability. 2, 7 (Jul. 2019), 2 621. DOI:https://doi.org/10.1038/s41893-019-0316-0. [2] Asensio, O.I. and Delmas, M.A. 2015. Nonprice incentives and energy conservation. Proceedings of the National Academy of Sciences. 112, 6 (Feb. 0 Rasch (medium effort & kWh) Rasch (high effort & kWh) 2015), E510–E515. DOI:https://doi.org/10.1073/pnas.1401880112. Rasch (low effort & kWh) Rating CF (med. EF&kWh) [3] Ekstrand, M.D. and Willemsen, M.C. 2016. Behaviorism is Not Enough: Rating CF (high EF & kWh) Rating CF (low EF & kWh) Better Recommendations Through Listening to Users. Proceedings of the 10th ACM Conference on Recommender Systems (New York, NY, USA, 2016), 221–224. [4] Elsweiler, D., Trattner, C. and Harvey, M. 2017. Exploiting Food Choice Top-10 RecLists per cluster - HighAbility Biases for Healthier Recipe Recommendation. Proceedings of the 40th International ACM SIGIR Conference on Research and Development in 6 Information Retrieval (New York, NY, USA, 2017), 575–584. [5] Kaiser, F.G. and Wilson, M. 2004. Goal-directed conservation behavior: the specific composition of a general performance. Personality and Individual Differences. 36, 7 (May 2004), 1531–1544. 4 DOI:https://doi.org/10.1016/j.paid.2003.06.003. [6] Kleppe, M., Lacroix, J., Ham, J. and Midden, C. 2015. The development of the ProMAs: a Probabilistic Medication Adherence scale. Patient Prefer Adherence. 9, (2015), 355–367. [7] McNee, S.M., Riedl, J. and Konstan, J.A. 2006. Being Accurate is Not 2 Enough: How Accuracy Metrics Have Hurt Recommender Systems. CHI ’06 Extended Abstracts on Human Factors in Computing Systems (New York, NY, USA, 2006), 1097–1101. [8] Schäfer, H., Hors-Fraile, S., Karumur, R.P., Calero Valdez, A., Said, A., 0 Rasch (medium EF & kWh) Rasch (high EF & kWh) Torkamaan, H., Ulmer, T. and Trattner, C. 2017. Towards Health (Aware) Rasch (low effort & kWh) Rating CF (med. EF & kWh) Recommender Systems. Proceedings of the 2017 International Conference on Rating CF (high EF & kWh) Rating CF (low EF & kWh) Digital Health (New York, NY, USA, 2017), 157–161. [9] Schäfer, H. and Willemsen, M.C. 2019. Rasch-based Tailored Goals for Nutrition Assistance Systems. Proceedings of the 24th International Conference on Intelligent User Interfaces (New York, NY, USA, 2019), 18–29. Figure 1. Rasch-based and Rating-based recommendation [10] Starke, A., Willemsen, M. and Snijders, C. 2017. Effective User Interface Designs to Increase Energy-efficient Behavior in a Rasch-based Energy sets for low-ability (top) and high-ability users (bottom), Recommender System. Proceedings of the Eleventh ACM Conference on divided among three clusters of energy-saving measures. Recommender Systems (New York, NY, USA, 2017), 65–73. [11] Starke, A.D. 2019. Supporting energy-efficient choices using Rasch-based recommender interfaces. (2019). Figure 1 shows the results for users with either a low (at the [12] Trattner, C. and Elsweiler, D. 2017. Food Recommender Systems: top) or high ability (at the bottom). Both graphs discern between Important Contributions, Challenges and Future Research Directions. three measure clusters (a combination of high/low effort and arXiv:1711.02760 [cs]. (Nov. 2017). [13] Trattner, C. and Elsweiler, D. 2019. What online data say about eating kWh savings). Figure 1 shows that Rating-based CF is not habits. Nature Sustainability. 2, 7 (Jul. 2019), 545. sensitive to changes in a user’s ability or motivation to engage in DOI:https://doi.org/10.1038/s41893-019-0329-8. 30