Food Recommender Systems: Important Contributions, Challenges and Future Research Directions

Introduction

• Diet-related illnesses account for 60% of deaths worldwide, while food recommendation remains comparatively understudied despite the importance and difficulty of food choices.Source: Introduction · p. 1

Developed Approaches

• Health-aware food recommenders are presented as a potential means of encouraging healthier nutritional choices, and the review also covers systems for healthy nutrition.Source: Introduction · p. 1, Developed Approaches · p. 3

Introduction

• Food recommendation is challenging because preferences are multifaceted, culturally determined, and context-dependent, while allergies, dietary preferences, ingredient availability, cooking knowledge, and equipment add constraints.Multiple names for foods and variations in ingredient preparation also complicate recommendation.Source: Introduction · p. 1
• The paper reviews food recommender algorithms, evaluation, and practical resources, then discusses challenges, lessons from prior research, and future directions.Source: Introduction · p. 1

Developed Approaches

• The authors selected 25 relevant papers covering recipes, meal plans, groceries, and menus, while excluding systems that recommend restaurants rather than food items.The literature overview is organized chronologically in Table 1.Source: Developed Approaches · pp. 1, 3, Table 1 · p. 2
• Table 1 characterizes contributions by algorithm, recommended item, feedback, context, dietary constraints, target users, and dataset.Its dimensions provide the structural basis for the remainder of the paper.Source: Table 1 · p. 2, Developed Approaches · p. 3
• The section reviews content-based, collaborative-filtering, and hybrid approaches, with contextual recommendation, group recommendation, and healthy-nutrition systems as further topics.Single-item recommendations, mostly recipes, are identified as the most prominent form.Source: Developed Approaches · p. 3
• Most reviewed work recommends recipes; less-studied food items include menus using restaurant reviews and groceries using online shopping data.The authors expect interest in other food items to grow as datasets become more available.Source: Developed Approaches · p. 3

Content-Based Methods (CB)

• Ingredient-based models used liked recipes to identify positive ingredient preferences, with later work also downweighting ingredients in disliked recipes (Freyne & Berkovsky, 2010; Freyne, Berkovsky & Smith, 2011; Harvey et al., 2013).Source: Content-Based Methods (CB) · p. 3
• Ingredient complement and substitution networks predicted preferences significantly better than approaches using ingredient lists, cooking methods, or styles as features (Teng et al., 2012).Source: Content-Based Methods (CB) · p. 3
• CNNs designed to extract visual food-image features outperformed baseline approaches, while low-level features such as brightness, colorfulness, and sharpness also helped predict preferences (Yang et al., 2015, 2017; Elsweiler, Trattner and Harvey, 2017).Source: Content-Based Methods (CB) · p. 3

Collaborative Filtering-Based Methods (CF)

• Pearson-correlation nearest-neighbor filtering on ratings performed worse than the content-based approach.Source: Collaborative Filtering-Based Methods… · p. 3
• SVD outperformed both the content and collaborative-filtering approaches proposed by Freyne & Berkovsky (2010), according to Harvey et al. (2013).Source: Collaborative Filtering-Based Methods… · p. 3
• The mobile interface example combines ratings with tags, which provide finer-grained feedback alongside binary or scaled ratings (Elahi et al., 2014; Ge, Elahi et al., 2015).Source: Figure 1 · p. 4, Collaborative Filtering-Based Methods… · p. 4
• Rating-and-tag matrix factorization significantly outperformed content-based and standard matrix-factorization baselines (Ge, Elahi et al., 2015).Source: Collaborative Filtering-Based Methods… · p. 4
• LDA and WRMF were the highest-performing collaborative-filtering approaches tested on a large Allrecipes.com dataset (Trattner and Elsweiler, 2017).Source: Collaborative Filtering-Based Methods… · p. 4

Hybrid Methods (Hybrid)

• Hybrid recipe recommenders combine collaborative filtering and content-based methods, while Harvey et al. (2013) found the best experimental performance by combining SVD with user and item biases.Freyne and Berkovsky (2010) combined user-based collaborative filtering with content-based methods; their group-focused follow-up switched among three strategies based on the ratio of a user’s rated items to all items.Source: Hybrid Methods (Hybrid) · p. 4

Context-Aware Approaches

• Exploratory analyses identified gender, time, hobbies, location, and food availability as relevant context in food recommendation.Studies used filtering to examine how recipes were rated, bookmarked, shared, or related to health statistics.Source: Context-Aware Approaches · p. 4
• Regression analyses found that recipe-preparation descriptions, nutrition, ingredient availability, and day of the week affect users’ ratings.Harvey et al. (2012) collected detailed context data about participants’ ratings of recipes as dinner choices for that day.Source: Context-Aware Approaches · p. 4
• The most important contextual variables and the best ways to account for them algorithmically remain unclear.Harvey et al. (2013) limited their algorithmic approaches to nutritional, user, and item biases despite studying numerous factors.Source: Context-Aware Approaches · p. 5
• Recipe recommendations for individual users typically perform more poorly on standard metrics than recommendations in other domains.The paper presents standard-technique experiments on a well-known recipe dataset in Table 2 to illustrate this challenge, but the supplied passage gives no values.Source: Context-Aware Approaches · p. 5

Group-Based Methods

• In a family study, Berkovsky and Freyne (2010) found personalized group recommendations worked best, but could not personalize recommendations for every user.They compared general popularity, aggregated user models, aggregated predictions, and personalized collaborative filtering strategies with real users.Source: Group-Based Methods · p. 5
• Elahi et al. (2014) used tags and role-based group preference elicitation, delivering meal recommendations to a designated leader or cook.The group utility score aggregated predictions using tags and ratings from all members.Source: Group-Based Methods · p. 5

Health-Aware Methods

• Direct nutrition-aware methods include Ge, Ricci and Massimo’s (2015) calorie-balance function and a linear framework balancing estimated user preference against nutritional error (Elsweiler, Harvey, Ludwig and Said, 2015).The linear framework’s λ parameter sets the relative weighting of rating and nutrition; λ = .5 gives equal weight.Source: Health-Aware Methods · p. 5
• Trattner and Elsweiler (2017) found post-filtering could balance accuracy and health, but the healthiest results had extremely poor accuracy and did not ensure healthy nutrition.The approach reweights item scores using WHO or inverse FSA health scores; LDA and WRMF were more suitable than MostPop and BPR.Source: Health-Aware Methods · pp. 5–6
• 4,025/6,400 cases (63%) yielded meal plans, with at least one plan for 58/64 profiles (91%) and all 100 personas.Automated plans searched breakfast and main-meal combinations against persona-specific nutritional targets; fixed portions struggled with high-calorie needs and narrow ingredient preferences. A manual interface also filtered recommendations so added meals kept plans within guidelines.Source: Health-Aware Methods · p. 6
• Elsweiler, Trattner and Harvey (2017) used predictive-model recipe replacements to nudge choices; participants were significantly more likely to choose recipes with much less fat.Visual and nutritional features were important to predicting choices.Source: Health-Aware Methods · p. 6
• Ingredient substitution has been proposed to improve recipes’ healthfulness but remains insufficiently evaluated nutritionally; Kusmierczyk, Trattner and Nørvåg (2016) explored recommendations using prior uploads and social, temporal, and geographic context.Source: Health-Aware Methods · p. 6
• Specialist diets remain poorly served and constraint-based methods are rare; Yang et al. (2017) applied dietary filters, while Ferguson, Darmon, Briend and Premachandra (2004) used linear programming for Malawian children.The latter approach sought to ensure children met expert-recommended nutritional intake.Source: Health-Aware Methods · pp. 6–7

User preference sources and scarcity

• Research mainly uses explicit ratings, bookmarks, or shares; implicit feedback is less common, and scarcity or cold-start has not been directly addressed.Examples of implicit feedback include recipe views and review sentiment; standard methods such as SVD have been applied to sparse data (Harvey et al., 2013; Trattner & Elsweiler, 2017).Source: Addressed Challenges and Problems · p. 7

Offline and online evaluation of recommendations

• Evaluation has been almost entirely offline, using naturalistic datasets or user studies; full-online evaluations have not been published to the authors’ knowledge.User studies have tested interfaces in semi-controlled and naturalistic settings, but these do not amount to full-online evaluations.Source: Addressed Challenges and Problems · p. 7

Beyond accuracy

• Accuracy remains unresolved; novelty and serendipity are appreciated but unstudied, and work on dietary diversity and preference–healthfulness trade-offs remains preliminary.Harvey et al. (2013) report user appreciation of novelty and serendipity; Elsweiler and Harvey (2015) acknowledge dietary diversity, while Elsweiler & Harvey (2015) and Trattner & Elsweiler (2017) provide preliminary work on the trade-off.Source: Addressed Challenges and Problems · pp. 7–8

Recommendation visualizations and explanations

• Recommendation visualizations and explanations have been superficial, and their evaluations have also been limited.Examples include Trattner and Elsweiler’s (2017) traffic-light system, Harvey and Elsweiler’s (2015) plan metadata, and explanations from Elahi et al. (2014); evaluations of these systems have been superficial.Source: Addressed Challenges and Problems · p. 8

Other common challenges

• The food domain reports no significant contributions on privacy, collaborative-recommender scalability or distribution, or robustness and attacks.Source: Addressed Challenges and Problems · p. 8

Challenges unique to food recommender systems

• Food-specific content work has normalized ingredients and quantities and studied the roles of context and visual features in food choices.Examples include Kusmierczyk et al. (2015) and Müller, Mika, Harvey & Elsweiler (2012) on ingredient processing, and Elsweiler, Trattner & Harvey (2017) and Yang et al. (2017) on visual features.Source: Addressed Challenges and Problems · p. 8
• Health research models nutrition and incorporates it into recommendations, but it remains unclear which recommendation method works most effectively.Work includes modeling user requirements, intake, and portion sizes; estimating recipe nutrition through ingredient matching or food-image analysis; and using nutrition-aware recommendations, meal plans, or algorithmic nudging (Schäfer et al., 2017; Gibney, Vorster & Kok, 2002; Straßburg, 2010; Zhang et al., 2011; Müller et al., 2012; Chokr & Elbassuoni, 2017; Ge, Elahi et al., 2015; Elsweiler & Harvey, 2015; Elsweiler, Trattner & Harvey, 2017).Source: Addressed Challenges and Problems · p. 8

• The section identifies datasets used to study food consumption patterns and evaluate algorithmic approaches.Source: Implementation Resources · p. 8
• It covers nutrition and health resources for implementing health-aware recommender systems.Source: Implementation Resources · p. 8
• It describes frameworks typically employed to build food recommender systems.Source: Implementation Resources · p. 8

• Food recommender research lacks standardized datasets, limiting the reliability and generalizability of published results.The field typically relies on proprietary, non-standardized collections, unlike domains with established benchmarks such as MovieLens.Source: Recipe, Meal plan, Menu and Grocery… · pp. 8–9

Recipes

• Recipe studies often use Allrecipes4 or Food.com5, but sharing restrictions leave few public profile datasets; Cookpad7 and Yummly8 provide licensed access, and Yummly offers a no-cost restricted-data API.Source: Recipe, Meal plan, Menu and Grocery… · p. 8
• MIT9 released over 1 million recipes with images and metadata, but lacks user profiles and interactions, limiting offline recommender evaluation.Source: Recipe, Meal plan, Menu and Grocery… · p. 9

Meal plans and restaurant menus

• No freely available meal-plan datasets are known, while Yelp10 has been used for menu recommendation but may require crawling because its terms omit data sharing.Source: Recipe, Meal plan, Menu and Grocery… · p. 9

Groceries

• Only one freely available grocery dataset is identified: Kaggle11’s 3 million Instacart purchases, with limited grocery metadata; other datasets are not public.Source: Recipe, Meal plan, Menu and Grocery… · p. 9

Nutrition & Health Resources

• Nutritional values can be estimated by mapping ingredients to USDA or BLS databases, with Müller et al. (2012) combining rules and learning to rank for German recipes.Edamam offers similar functionality for English and Spanish recipes; the Müller framework is available from its authors, but using it requires a BLS license.Source: Nutrition & Health Resources · p. 9
• Ingredient normalization must reconcile aliases, preparation, and units; NLP or Spoonacular can unify ingredient extraction, while food word lists and Wikipedia can support normalization.Standard NLP techniques include stop-word removal, conjunction splitting, and string matching; cooking methods may also affect nutritional values.Source: Nutrition & Health Resources · pp. 9–10
• FSA and WHO guidelines use nutrient ranges based on a 2,000 kcal diet; FSA traffic-light scores can be summed to yield a 4-to-12 healthiness scale, while HEI has not been used in food recommender systems.FSA labels use red, amber, and green; Sacks, Rayner and Swinburn (2009) assign these values of 3, 2, and 1, respectively.Source: Nutrition & Health Resources · pp. 9–10, Table 4 · p. 10, Table 5 · p. 10
• The foodsubs thesaurus can suggest healthier ingredient substitutes and may also help accommodate allergies or intolerances (Achananuparp & Weber, 2016).Source: Nutrition & Health Resources · p. 10
• CDC state and county diabetes and obesity data may help tailor food recommender systems to regional differences in eating patterns (Trattner, Parra & Elsweiler, 2017; Said & Bellogín, 2014).Source: Nutrition & Health Resources · p. 10

Food Recommender System Frameworks

• Food recommender research mostly relies on researcher-built software, and the lack of a shared open-access framework makes progress and reproduction or validation of prior findings difficult.Source: Food Recommender System Frameworks · p. 10
• LibRec provides standard recommendation algorithms for rating prediction and item ranking, and Trattner & Elsweiler (2017) used its pre- and post-filtering to re-rank recipes by healthiness.Source: Food Recommender System Frameworks · pp. 10–11
• The framework can be extended to group recipe recommendations and personalized meal plans, and the authors offer to share their code upon request.Source: Food Recommender System Frameworks · p. 11

• Early systems used CBR planning to turn queries into meal plans or recipes, then simple content-based and collaborative filtering variants appeared.Source: Historical Evolution and Versions of… · p. 11
• The first directly comparable modern food recommenders used standard algorithms such as UserKNN with recipe datasets and reliable evaluation.Source: Historical Evolution and Versions of… · p. 11
• Later approaches derived recipe similarity for healthful meals, used knowledge bases (El-Dosuky et al., 2012), or linked recipes via tag-derived knowledge graphs.Ingredient networks were proposed for recommendations, and a model accounted for food-selection biases (Teng et al., 2012).Source: Historical Evolution and Versions of… · p. 11
• Yang et al. (2017) introduced a constraint-based mobile recommender that learned preferences from food images, unlike earlier approaches based on ratings or tags (Ge, Elahi et al., 2015).Source: Historical Evolution and Versions of… · p. 11
• Behavioral work found preference for unhealthy recipes (Trattner & Elsweiler, 2017), identified consumption differences by hobbies and gender (Trattner, Rokicki & Herder, 2017; Rokicki et al., 2016), and first addressed health-aware recipe recommendation.Elsweiler, Trattner & Harvey (2017) investigated the extent to which food recommenders can nudge people toward healthier choices.Source: Historical Evolution and Versions of… · p. 11

• Historical-data simulations are most common, typically using an 80/20 train-test split, though studies also use k-fold validation, leave-one-out, and varied data sources.Source: Evaluation: Metrics and Methodologies · p. 11
• Evaluation has shifted from rating-prediction errors such as MAE and RSME toward ranked-list metrics such as Recall, Precision, MAP, and NDCG (Cheng et al., 2017; Trattner & Elsweiler, 2017; Yang et al., 2017).Source: Evaluation: Metrics and Methodologies · p. 11
• Health-oriented evaluation can use governmental or organizational guidelines and ∆FSA and ∆WHO to measure healthfulness differences between users’ positively rated and predicted items (Trattner & Elsweiler, 2017).These delta measures indicate whether recommendations are more or less healthy than items users rated positively.Source: Evaluation: Metrics and Methodologies · p. 12
• Online evaluations are rare: one live meal-planner study compared personalized and non-personalized methods with over 5,000 users, while Yang et al. (2017) tested recommendations with 60 participants.Source: Evaluation: Metrics and Methodologies · p. 12
• Three online studies with 107–138 participants each examined whether recipe information could nudge healthier choices and found models could predict preferences and influence choices (Elsweiler, Trattner & Harvey, 2017).Source: Evaluation: Metrics and Methodologies · p. 12
• Food recommenders lack specialized offline and online evaluation protocols; standard metrics assess prediction accuracy and diversity, with WHO and FSA scores as health-oriented exceptions (Trattner & Elsweiler, 2017).Source: Evaluation: Metrics and Methodologies · p. 12

Lessons Learned and Future Directions

• Food recommendation is relatively under-researched, and existing methods predict food preferences less effectively than methods in other domains.The paper argues the domain is important for individuals and society, but that current performance remains poorer than elsewhere.Source: Lessons Learned and Future Directions · p. 12
• Standard preference-learning methods produce relatively unsatisfactory performance across ratings, tags, and comments, so new methods are needed.Image-based preference learning using embeddings from deep neural networks has shown promise.Source: Lessons Learned and Future Directions · pp. 12–13
• Recommendation models should incorporate important context variables, potentially captured through different sensors.The literature highlights context variables as important for preference prediction.Source: Lessons Learned and Future Directions · p. 13
• Food recommenders should more concretely account for social situations and how shared culinary experiences influence food choices.The paper calls for technological systems to reflect the pervasiveness of social culinary experiences.Source: Lessons Learned and Future Directions · p. 13
• Nutritional-health recommenders use nutrition-aware algorithms, meal plans, and nudging, but these proposals remain preliminary and no best approach is established.The paper identifies nutritional health as a prominent research focus for societal and socio-economic reasons.Source: Lessons Learned and Future Directions · p. 13
• The community should develop standard datasets and baseline methods, and conduct more online studies of food recommenders in naturalistic settings.Evaluation in the literature has mostly relied on offline studies using proprietary collections.Source: Lessons Learned and Future Directions · p. 13