Healthy Menus Recommendation: Optimizing the Use of the Pantry Jefferson Caldeira Ricardo S. Oliveira Fed. Uni. of Campina Grande, Brazil Fed. Uni. of Campina Grande, Brazil jeffersonemanuel@copin.ufcg.edu.br ricardooliveira@copin.ufcg.edu.br Leandro Marinho Christoph Trattner Fed. Uni. of Campina Grande, Brazil University of Bergen, Norway lbmarinho@computacao.ufcg.edu.br christoph.trattner@uib.no ABSTRACT rush of everyday life is often present. By menu, we mean a We are often unable to plan menus ahead, thus making poor set of meals, where each meal is comprised by a set of recipes and unhealthy choices of meals. Besides healthy, one may and each recipe is comprised by a set of ingredients. In this want menus in which ingredients harmonize and cover well paper, we will focus on lunch menus, leaving other types of the available ingredients in the pantry. In this paper, we pro- menus for future work. pose a novel multi-objective-based recommender of menus Besides fulfilling healthy nutritional standards, other prop- that features an optimal balance between nutritional aspects, erties should be considered before planning a menu, such harmony and coverage of available ingredients. We conduct as the harmony and easy availability of ingredients. In this experiments on real-world and synthetic datasets and show paper, we introduce a novel lunch menu recommendation that our approach achieves the desired levels of nutrients, approach that considers all these properties simultaneously. harmonization and coverage of ingredients. Our algorithm receives as inputs the ingredients available at the user’s pantry and the number of portions desired, CCS CONCEPTS and recommends a lunch menu composed of a set of meals. • Information systems → Recommender systems; • Ap- The number of meals in the menu is decided automatically plied computing → Consumer health; Health informatics; based on the availability of ingredients in the pantry. If a recommended menu has 7 meals, for example, the user could KEYWORDS choose one different lunch meal for each day of the week. Recommender systems; meal recommendations; nutrients Each meal is composed of main dish, three side dishes di- optimization; multiobjective optimization vided into rice, beans, and pasta, salad, beverage, and dessert. An example of such a meal is roast chicken (main dish), garlic ACM Reference Format: rice, black beans, spaghetti (side dishes), broccoli salad (salad), Jefferson Caldeira, Ricardo S. Oliveira, Leandro Marinho, and Christoph orange juice (beverage), and gelatin (dessert)1 . We have cho- Trattner. 2018. Healthy Menus Recommendation: Optimizing the sen a meal setup particularly suited to the Brazilian food Use of the Pantry. In Proceedings of the Third International Work- culture, where three of the authors reside, although it could shop on Health Recommender Systems co-located with Twelfth ACM Conference on Recommender Systems (HealthRecSys’18), Vancouver, be easily reconfigured to other food cultures. BC, Canada, October 6, 2018 , 6 pages. We cast this as a multi-objective optimization problem, where standard nutritional indexes, harmonization and cov- 1 INTRODUCTION erage of ingredients in the pantry are formulated as (possibly conflicting) objective functions. We use the Non-dominated There is a growing recognition that healthy food directly in- Sorting Genetic Algorithm II (NSGA II) [7], which besides fluences quality of life, by providing a sense of well being and providing guarantees of convergence, feature diversity of happiness. However, eating healthily remains a challenge for recipes as an intrinsic property of the solution. We conduct many people. Among the possible reasons, the inability to experiments on real-world and synthetic data, and show that conciliate the planning of healthy and tasty menus with the our approach is able to achieve an optimal balance between the desired level of nutrients, harmonization and coverage HealthRecSys’18, October 6, 2018, Vancouver, BC, Canada of ingredients. © 2018 Copyright for the individual papers remains with the authors. Copy- ing permitted for private and academic purposes. This volume is published 1 In this work we do not consider vegetarian meals, so a menu always and copyrighted by its editors. contains some kind of meat. HealthRecSys’18, October 6, 2018, Vancouver, BC, Canada Caldeira et al. Table 1: Notation used to describe our approach. Harmony. For measuring the harmony of a meal, we design Sym. Description an objective function where two ingredients are considered N set of nutritional components {proteins, carbohydrates, total fat }. to harmonize well if they co-occur often in different recipes R set of recipes. in the dataset. For each pair of ingredients in a meal, we C set of recipes categories, i.e., main dish, side dish 1, side dish 2, side dish 3, salad, beverage and dessert. compute the relative co-occurrence frequency of these in- M set of meals. P set of pantries. gredients considering all recipes in which they appear as m ∈M set of recipes where each recipe belongs to a different category of ingredients. Eq. 3, defined in R ≥0 , formalizes this idea. C. дn (r ) nutritional value of recipe r ∈ R regarding n ∈ N . f n (m) nutritional value of meal m ∈ M regarding n ∈ N , i.e., Õ Ri ∩ R j harm(m) = (3) r ∈m дn (r ). Ri ∪ R j Í I set of ingredients, e.g., rice and tomato. i, j ∈Im ,i,j Im set of ingredients present in meal m ∈ M . Ip set of ingredients present in pantry p ∈ P . where a value of 0 means that the ingredients of m do not Ri set of recipes containing ingredient i ∈ I . Vn ⊆ R≥0 range [min, max] of reference values for n ∈ N . harmonize at all. maxVn , minVn maximum and minimum values of the interval Pn resp. maxn maximum possible value of the nutritional component n ∈ N Coverage. We seek to recommend menus that use, as much found in the set of possible meals. qm (i) required quantity in grams of ingredient i ∈ I for preparing meal as possible, the available ingredients in the pantry. For that, m ∈ M , e.g, it is required 200д of rice for preparing a meal having we design a coverage function as the ratio between the avail- garlic rice as side dish. qp (i) available quantity in grams of ingredient i ∈ I in pantry p ∈ P . able ingredients in the pantry and the required ingredients for composing the meal. Eq. 4, defined in [0, 1], formalizes 2 PROBLEM FORMALIZATION this idea. For recommending menus featuring healthy nutritional val- i ∈Im qp (i) , i ∈Im qm (i) Í  Í  min ues, good harmonization and coverage of ingredients in the cov(m, t) = (4) i ∈Im qm (i) × t Í user’s pantry, we first design specific objective functions for each one of these properties. Please refer to Table 1 for where t is the number of portions required. understanding the symbols used in this section. For example, suppose that a certain meal requires 200д of rice for one person, i.e., qm (rice) = 200. If we consider 3 Nutrition. A meal m ∈ M is composed of seven recipes, portions (t = 3), we will require 600д of rice. If the pantry has each one belonging to a different category of C. For measur- available 600д of rice exactly, i.e., qp (rice) = 600, then the ing if a meal complies to daily recommended lunch nutri- coverage is maximum with a value of 1. If the meal requires tional values, regarding the nutritional component n ∈ N , we less than what is available in the pantry, the coverage should define distn (m) ∈ [0, 1] for computing the distance between also be maximum since we found all required ingredients in the nutritional value of a meal and the range of nutritional the pantry (that is why the min in the numerator of Eq. 4). reference values Vn . More formally, distn (m) is defined as: If, however, the meal requires more than what is available, | fn (m) − minVn | + | fn (m) − maxVn | − (maxVn − minVn ) the function returns a value less than 1. So we seek meals maxn that maximize this function. (1) If the nutritional values of the meal fall inside Vn , the Problem Statement. Given a set of ingredients Ip available function returns 0, otherwise it returns a value higher than at some pantry p ∈ P and the number t of portions required, 0. Values close to 1 mean that m ∈ M has a nutritional our goal is to find the lunch meals that maximize all the value close to the highest possible meal nutritional value aforementioned functions simultaneously, i.e., ! ! in the dataset. In this work, we have used reference values k Õ provided by the Ministry of Health of Brazil (cf. Section 4). arg max nutn (m) , harm(m), cov(m, t) (5) m For example, suppose that a given meal m ∈ M has a protein n ∈N value of 120д, i.e., f prot (m) = 120, reference protein values in where k is the number of meals returned, calculated automat- Vprot = [100, 150], and the maximum possible protein value ically by our algorithm (cf. Section 5). Notice that every time of maxprot = 500. Applying distn (m) we have a meal that maximizes Eq. 5 is selected, the pantry needs to |120 − 100| + |120 − 150| − (150 − 100) be updated accordingly and the coverage of the subsequent =0 meals have to take these updated values into consideration. 500 meaning that m complies with the protein reference values. 3 RELATED WORK For later convenience, we will seek the meal that maxi- Several related works have been proposed with the aim of mizes the inverse of the distance: recommending food to people. Trattner and Elsweiler [22] nutn (m) = 1 − distn (m) (2) provides a good overview in this direction showing advances Healthy Menus Recommendation HealthRecSys’18, October 6, 2018, Vancouver, BC, Canada in recommender technology in the context of recipes, gro- as we do. Moreover, our approach takes into consideration a ceries or meals. larger number of criteria in comparison to previous methods. One of the earliest examples in this area are the works In all, our contributions are summarized as follows: of Hammond [15] and Hinrichs [17] where case-based rea- • A novel method to recommend lunch menus consid- soning methods to recommend meal plans and recipes to ering, at the same time, reference nutritional values, people are introduced. More advanced algorithms employ- harmony of ingredients, and coverage of pantry. ing content-based filtering and early stages of collaborative • Tailor designed objective functions for each property filtering include the works of Lawrence et al. [18], proposing under consideration; a method to recommend groceries, and the work of Aberg • The recommendation approach features easy to ex- [1], proposing a recommender method to nourish elderly plain recommendations; people properly. • Experiments showing that the recommended menu Freyne and Berkovsky [11] introduced further advances achieves the expected values of the desired properties. by employing user-based K-NN collaborative filtering to recommend recipes. Subsequently, more advanced methods 4 NUTRITIONAL REFERENCE VALUES and algorithms emerged for tackling different problems and The World Health Organization (WHO)2 is responsible for, aspects related to food recommendation. A good example among other things, setting norms and standards and assess- in this direction is the work of Teng et al. [21] proposing a ing health trends world wide. It provides up-to-date refer- novel recipe recommender method based on ingredient net- ences about healthy diets, which are used by many govern- works. Other relevant studies include the work of Berkovsky ments and institutions around the world for the definition and Freyne [4] proposing a method to recommend meals to of their own health policies. groups of people; Harvey et al. [16] proposing a model that The Ministry of Health of Brazil, for example, produced accounts for food selection biases; Ge et al. [13] proposing a technical reports Ministério da Saúde [19] and a program method that leverages tags and latent factors to recommend called PAT - Programa de Alimentação do Trabalhador (Work- recipes; Yang et al. [25] proposing the first constraint-based ers’ Nutrition Program) [8], containing nutritional reference (with different types of diets) recommender, and the more values for healthy meals based on WHO, which we adopted recent work of Trattner et al. [24] proposing a novel method in this work. In particular, we have adopted, considering an to recommend recipes to people in a cold-start scenario. adult, a recommended energy intake of 2,000 kcal3 , which Other recent relevant works include Trattner and Elsweiler results in a range between 600 and 800 kcal for the lunch [23] showing the extent to which current recommendation meal. The reference values we have used, considering all algorithms are suitable for recommending healthy recipes. nutritional components, are summarized in Table 2. They were also the first to employ the WHO standards to recommend healthy recipes and meal plans. Also of relevance are [12, 14] or [9] and [10] proposing Table 2: Nutritional components reference values used a novel method to bring the ‘healthiness’ aspect into meal in this paper. Minimum Maximum plans. In this direction, it is worth mentioning the works of Proteins 60g 120g Chifu et al. [5], proposing a solution using Particle Swarm Carbohydrates 330g 600g Optimization to build healthy daily menu recommendations Total fat 90g 240g for elderly people; Agapito et al. [2] proposing a recom- mender system with focus on patients with chronic diseases; Cholissodin and Dewi [6] taking into account the family bud- 5 MULTI-OBJECTIVE MENU RECOMMENDATION get; and the work of Seljak [20] proposing a multi-objective Our approach receives as input: approach for developing nutritionally and gastronomically • A set of recipes, categorized as: main dish (meat, chicken, adequate menus. pork, fish, etc.), three side dish categories (rice, beans and pasta), salad, beverage and dessert; Summary & Contributions. The related works reveal sev- • A shopping basket, containing products normally used eral solutions available to tackle the food recommendation as ingredients for food preparation, as well as its quan- problem. Differently from us that recommend several meals tities, representing the user’s pantry; grouped in a menu, most of these solutions focus on rec- • A number of portions corresponding to the number of ommending recipes. Part of these solutions are concerned persons for whom the meals will be prepared. in recommending healthy food to people. Interestingly, we are not aware of any work that takes into consideration the 2 http://www.who.int/ ingredients that the user has available to prepare her food, 3 This is in line with the WHO reference values. HealthRecSys’18, October 6, 2018, Vancouver, BC, Canada Caldeira et al. The decision of using a multi-objective approach is sup- Dataset User's Process ported by the Pareto dominance concept, which is useful to Recipes shopping Manual Input collected compare different solutions across multiple objectives. For baskets Set of Items two candidate solutions, we say that one dominates the other Filtering if one solution is better than the other in at least one objec- Identify Identify Identify Identify Re- and ingredients quantities ingredients quantities sampling tive, when there is a tie in all others. For example, consider grouping the objectives related to harmony and protein levels. Sup- pose that two meals are being compared, the first one with Number of Recipes with ingredients User pantries and amounts for one harmony and protein levels of 0.9 and 80д and the second portions portion one with 0.85 and 110д respectively. In this case, the first NSGA-II meal has a superior value of harmony, but there is a tie in Nutritional protein since both meals fall into the healthy reference range Coverage Quality Harmony for protein (60д to 120д). Thus, the first meal dominates the second with respect to these objectives. The set of all non- Evaluate Select meals Recommended Menu dominated solutions is called Pareto Front (or Pareto-optimal thresholds according to Meal 1 Meal 3 solutions) which represents the set of best possible solutions pantry Meal 2 ... with respect to the objectives considered. In our case, each candidate solution is a meal. In order to Figure 1: Pipeline of our recommendation approach. find the Pareto Front, it is necessary to compare each meal to every other meal in terms of the objective functions of inter- ingredients needed are available in the pantry. In this work, est, which leads to a combinatorial explosion problem. We we used 50% as this threshold, i.e., meals are recommended employ the NSGA-II algorithm [7] for solving this problem. only if at least 50% of the required ingredients are available The reason for choosing NSGA-II is twofold: (i) it converges in the pantry. Notice that the selection of a meal causes a to Pareto-optimal solutions at lower complexity than a brute- reduction of ingredients in the pantry, influencing the selec- force exhaustive search and (ii) it seeks to create diverse tion of the next meal. This selection is performed in a greedy Pareto-optimal solutions. Concerning (i), the complexity of manner, following the ranking provided by NSGA-II. Fig- NSGA II is O(MN 2 ) where M is the number of objective ure 1 summarizes the whole process of our recommendation functions and N the number of meals. Concerning (ii), this approach. is particularly interesting because it may end up favoring meals that are diverse in terms of the recipes and ingredients 6 EVALUATION used. In this section we present the evaluation of our approach. In the first iteration, NSGA II randomly selects a parent All code for the evaluation is publicly available online4 . generation of individuals (meals in this case). The size of this initial population is preserved across iterations. At each iter- Data Collection and Preparation. The set of recipes used ation, NSGA-II generates an offspring population through in this paper was collected from TudoGostoso5 , a Brazilian mutation (replacing one of the recipes) and crossing over website of food recipes similar to Allrecipes.com6 . This is one (switching recipes between meals). The individuals are or- of the most popular food websites in Brazil7 . A web crawler dered by domination ranges in a process called Fast Non- was implemented to collect the recipes, being executed from dominated Sorting, in which each range contains individuals 07/19/2017 to 08/28/2017. that do not dominate each other, but dominate the individu- In total, 12, 930 recipes were collected. Recipes about soups, als in the next range. alcoholic drinks, breads, snacks, etc., were discarded. We also If the number of individuals exceeds the population size, collapsed the categories meat, chicken and fish, into main some individuals in the last domination range are selected dish. In order to determine the nutritional information of a in a process called Crowding Distance Sorting. This is done in recipe, we first extracted the ingredients and their quantities order to spread the solutions along the Pareto Front, instead present in the HTML. We then passed these ingredients to of concentrating solutions around similar objective values. Tabela de Alimentos8 , a Brazilian website that receives an This process is particularly useful in our context, since it can ingredient name (in Portuguese) and its quantity, and returns improve diversity of the meals along the iterations. 4 https://github.com/JeffersonEmanuel/healthy-menus-recommendation After a certain number of iterations, NSGA-II will yield 5 http://www.tudogostoso.com.br/ a meal population that is Pareto-optimal. The final recom- 6 https://www.allrecipes.com/ mendation is formed by a set of meals extracted from this 7 See http://alexa.com/ 8 http://www.tabeladealimentos.com.br population, subject to the condition that a percentage of the Healthy Menus Recommendation HealthRecSys’18, October 6, 2018, Vancouver, BC, Canada (a) Nutritional results (b) Coverage and harmony results Figure 2: Experimental results: our approach against meals formed by randomly chosen recipes, considering one portion. nutritional information such as calories, carbohydrates, pro- Evaluation Protocol. The number of portions, represent- tein, total fat, fiber and sodium, about that ingredient. We ing the number of persons who will consume the meals, is wrote a script for automatizing this process. In this work, another input necessary to the experiments. We run the ex- the relevant information are carbohydrates, protein and fat. periments with the number of portions varying from one For many reasons, such as typos and synonyms, ingredi- to four. Each recipe evaluated has its ingredients’ quantities ent names may not be found in Tabela de Alimentos. Thus, multiplied by the number of portions, in order to determine recipes containing any ingredient for which nutritional infor- the availability of ingredients in the pantry. mation were not found, were discarded. After this filtering, As baseline, we have used a random approach that will 741 recipes remained, divided into the seven categories used keep selecting random meals while at least 50% of the re- to compose the meal, as described in Table 3. quired ingredients are available in the pantry. This will serve to confirm that that our solution is not by chance. Table 3: Number of recipes in each category. Meat Rice Beans Pasta 255 69 48 84 Salad Beverage Dessert TOTAL Results. The experimental results show that, for a group 90 75 120 741 of 1, 000 pantries, our approach can recommend lunch meals that fit the daily nutritional requirements in what concerns The dataset of shopping baskets, used to represent users’ proteins, carbohydrates and total fat, at the same time that pantries, contains 28 baskets. This dataset was collected from provides good harmony and a good use of ingredients in the fellow graduate students and contain both shops made con- pantry. sidering an entire family and people living alone. Each shop- Figure 2 shows the results. We show the box plots for 1 ping basket is filtered in order to contain only food items portion, that is, the meals are intended for one person, but (ingredients) and their respective quantities in the base unit the results are similar considering 2 to 4 portions. The blue of measure. horizontal lines in the left hand side of Figure 2 represent Personalization is achieved by recommending meals based the recommended range for each nutrient. on the ingredients available in user’s pantry. Due to the First, in comparison to the random baseline, Wilcoxon low number of pantries collected, a re-sampling was made tests showed that the distributions are significantly different in order to achieve the number of 1, 000 shopping baskets. for every tested objective function, with 95% confidence. Sec- This was performed as follows. Two baskets are randomly ond, regarding the nutritional components, most of the rec- sampled, where one of these baskets will receive a random ommended meals fall inside the recommended range, while number of new ingredients from the other basket (without the random approach has an erratic behavior as expected. repetition) where the quantities of each new ingredient are Finally, the meals recommended by our approach present multiplied by a number in the range [0.5, 1.5]. So the quan- better values of harmony and coverage than random, as ex- tities are shrinked, expanded or unchanged. We repeat this pected. process until 1, 000 baskets are produced. HealthRecSys’18, October 6, 2018, Vancouver, BC, Canada Caldeira et al. 7 CONCLUSIONS [6] Imam Cholissodin and Ratih Kartika Dewi. 2017. Optimization of Healthy Diet Menu Variation using PSO-SA. Journal of Information In this paper, we proposed a new approach for recommend- Technology and Computer Science 2, 1 (2017), 28–40. ing lunch menus, in which a menu is defined as a set of meals. [7] Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and TAMT Meyarivan. We cast the problem as multi-objective optimization prob- 2002. A fast and elitist multiobjective genetic algorithm: NSGA-II. IEEE lem where healthy nutrients, harmonization and coverage transactions on evolutionary computation 6, 2 (2002), 182–197. of ingredients in the pantry are considered simultaneously. [8] Programa de Alimentação do Trabalhador. 2006. Portaria interministe- rial nº. 66, de 25 de agosto de 2006. Publicada no DOU de 28 (2006). To the best of our knowledge this is the first approach to [9] David Elsweiler, Morgan Harvey, Bernd Ludwig, and Alan Said. 2015. consider this multitude of meals properties. We have used Bringing the" healthy" into Food Recommenders.. In DMRS. 33–36. the NSGA II algorithm, a state-of-art multi-objective opti- [10] David Elsweiler, Christoph Trattner, and Morgan Harvey. 2017. Ex- mization solver, that also features diversity of recipes and ploiting Food Choice Biases for Healthier Recipe Recommendation. In ingredients as an intrinsic property of the solution. Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR ’17). ACM, 575–584. Although our setup was tuned to fit Brazilian food culture [11] Jill Freyne and Shlomo Berkovsky. 2010. Intelligent Food Planning: and health standards, it can easily accommodate any food cul- Personalized Recipe Recommendation. In Proceedings of the 15th Int. ture and health reference values. We conducted experiments Conference on Intelligent User Interfaces (IUI ’10). ACM, 321–324. on real and synthetic data that confirms the soundness and [12] B Gaál, I Vassányi, and G Kozmann. 2005. A novel artificial intelligence quality of our approach. As future works, we intend to com- method for weekly dietary menu planning. Methods Archive 44, 5 (2005), 655–664. pare our approach with state-of-the-art food recommender [13] Mouzhi Ge, Mehdi Elahi, Ignacio Fernaández-Tobías, Francesco Ricci, systems with a related purpose, specifically the works of and David Massimo. 2015. Using Tags and Latent Factors in a Food Rec- Ahn et al. [3] and Cholissodin and Dewi [6]. Other purpose ommender System. In Proceedings of the 5th International Conference is to test the method with other recipe collections such as on Digital Health 2015 (DH ’15). ACM, New York, NY, USA, 105–112. Allrecipes.com in the US or Kochbar.de9 , one of the largest [14] Mouzhi Ge, Francesco Ricci, and David Massimo. 2015. Health-aware food recommender system. In Proceedings of the 9th ACM Conference recipe platforms in Europe. Also, we intend to employ dif- on Recommender Systems. ACM, 333–334. ferent standards as set by the World Health Organization [15] Kristian J Hammond. 1986. CHEF: A model of case-based planning. In (WHO) and Food Standard Agency (FSA) in the UK. Finally, Proceedings of AAAI. 267–271. we plan to conduct user studies in order to investigate quali- [16] Morgan Harvey, Bernd Ludwig, and David Elsweiler. 2012. Learning tative aspects of our approach. user tastes: a first step to generating healthy meal plans. In First int. workshop on recommendation technologies for lifestyle change. 18. [17] Thomas R Hinrichs. 1989. Strategies for adaptation and recovery ACKNOWLEDGMENTS in a design problem solver. In Proceedings of the 2nd Workshop on This work has been supported by the project ATMOSPHERE Case-Based Reasoning. 115–118. (atmosphere-eubrazil.eu), funded by the Brazilian Ministry of [18] Richard D Lawrence, George S Almasi, Vladimir Kotlyar, Marisa Science, Technology and Innovation (Project 51119 - MCTI/RNP Viveros, and Sastry S Duri. 2001. Personalization of supermarket product recommendations. In Applications of Data Mining to Electronic 4th Coordinated Call) and by the European Commission un- Commerce. Springer, 11–32. der the Cooperation Programme, Horizon 2020 grant agree- [19] Departamento de Atenção Básica. Secretaria de Atenção à Saúde Min- ment no 777154. istério da Saúde. 2008. Guia alimentar para a população brasileira: promovendo a alimentação saudável. (2008). REFERENCES [20] Barbara Koroušić Seljak. 2006. Dietary menu planning using an evolu- [1] Johan Aberg. 2006. Dealing with Malnutrition: A Meal Planning Sys- tionary method. In 2006 International Conference on Intelligent Engi- tem for Elderly.. In AAAI Spring Symposium: Argumentation for Con- neering Systems. IEEE, 108–113. sumers of Healthcare. 1–7. [21] Chun-Yuen Teng, Yu-Ru Lin, and Lada A Adamic. 2012. Recipe recom- [2] Giuseppe Agapito, Mariadelina Simeoni, Barbara Calabrese, Giorgio mendation using ingredient networks. In Proceedings of the 4th Annual Fuiano, and Mario Cannataro. 2017. DIETOS: a recommender system ACM Web Science Conference. ACM, 298–307. for health profiling and diet management in chronic diseases. In Second [22] Christoph Trattner and David Elsweiler. 2017. Food Recommender International Workshop on Health Recommender Systems. 32. Systems: Important Contributions, Challenges and Future Research [3] Yong-Yeol Ahn, Sebastian E Ahnert, James P Bagrow, and Albert-László Directions. arXiv preprint arXiv:1711.02760 (2017). Barabási. 2011. Flavor network and the principles of food pairing. [23] Christoph Trattner and David Elsweiler. 2017. Investigating the health- Scientific reports 1 (2011), 196. iness of internet-sourced recipes: implications for meal planning and [4] Shlomo Berkovsky and Jill Freyne. 2010. Group-based recipe recom- recommender systems. In Proceedings of the 26th International Confer- mendations: analysis of data aggregation strategies. In Proceedings of ence on World Wide Web. 489–498. the fourth ACM conference on Recommender systems. ACM, 111–118. [24] Christoph Trattner, Dominik Moesslang, and David Elsweiler. 2018. [5] V Chifu, R Bonta, E St Chifu, I Salomie, and D Moldovan. 2017. Particle On the predictability of the popularity of online recipes. EPJ Data Swarm Optimization Based Method for Personalized Menu Recom- Science 7, 1 (2018), 20. mendations. In International Conference on Advancements of Medicine [25] Longqi Yang, Cheng-Kang Hsieh, Hongjian Yang, John P Pollak, Nicola and Health Care through Technology 2016. Springer, 232–237. Dell, Serge Belongie, Curtis Cole, and Deborah Estrin. 2017. Yum-Me: A Personalized Nutrient-Based Meal Recommender System. ACM 9 https://www.kochbar.de/ Transactions on Information Systems (TOIS) 36, 1 (2017), 7.