This paper presents a personalized recommendation system that suggests recipes to users based on their health history and similar users' preferences. Specifically, the system utilizes collaborative filtering to determine other users with similar dietary preferences and exploits this information to identify suitable recipes for an individual. The system is able to handle a wide range of health constraints, preferences, and specific diet plans, such as low-carb or vegetarian. We demonstrate the usability of the system through a series of experiments on a large real-world data set of recipes. The results indicate that our system is able to provide highly personalized and accurate recommendations.
1. Introduction
ing increasingly important across many industries,
using data mining and machine learning to analyze large
amounts of data and make personalized
recommendations to users [
On the other hand, the prevalence of chronic health tion with morbidity and mortality is well-documented[13, 14]. To reduce the risk of developing or exacerbating these conditions - as could be caused by poor dietary Published in the Workshop Proceedings of the EDBT/ICDT 2023 Joint overall coverage, and explainability of the results. The conditions has risen to alarming levels, and their associa- implementation of a python framework called SHARE. survey respondents also indicated their preferred method. dation process, Tian et al. [26] proposed a recipe RS that The results indicate that the system provides a wide vari- incorporates user history behavior and user feedback ety of highly relevant personalized recipes to users and such as the ratings toward recipes, which make the recexcellent justifications ommendation accounts for user interest and preferences.
The rest of the paper is as follows: Section II presents Similarly, Pessemier et al. [27] present a food recomrelated works. Section III presents our framework. Sec- mendation strategy for patients in a care facility that tion IV evaluates the SHARE framework, and finally, utilizes explicit ratings for menu items, implicit feedback Section V ends with a conclusion and future work. based on the patient’s eating behavior and the amount of food that was eaten, and inferred preferences for the ingredients of the menu items. 2. Related Work In addition, there are other approaches such as Hybrid methods, which combine two or more diferent recomThere has been a significant amount of previous work on mendation techniques, for example, Gaudani and Hetal recommendation systems in the field of food and recipe [28] proposed a hybrid approach that combines CB and recommendation based on user health. CF algorithms to recommend recipes.
One line of research has focused on developing content- SHARE is closely related to these previous studies and based RS, which recommends items based on their fea- builds upon them by combining both content-based and tures and characteristics. For example, Herlocker et collaborative filtering methods to develop a personalal.[15] present a content-based RS for music, which rec- ized recipe recommender system that takes into account ommends songs based on the audio features of the songs users’ preferences and health history. However, SHARE and the user’s previous listening history. Similarly, Freyne difers from the above works, because we also combine et al.[16, 17] propose a CB recipe recommender that rec- a knowledge-based component. Also, in content-based ommends recipes based on the ingredients and cooking ifltering where we use this component to analyze the numethods of the recipes. tritional information of the recipes to recommend recipes
Another line of research has focused on collaborative that are suitable for the user’s chronic disease. Overall acifltering methods, which recommend items based on the cording to our research, there is no other hybrid approach preferences of similar users. For example, Resnick et al. that combines collaborative filtering, content-based filter[18] present a CF method for a movie recommendation, ing, and knowledge-based methods where the CB comwhich recommends movies to users based on the ratings ponent extracts tags from recipe descriptions including of similar users. Likewise, Freyne et al. [19], propose a nutritional information and other relevant characteriscollaborative filtering method for recipe recommenda- tics. tion, which recommends recipes to users based on the ratings of similar users.
In the health domain, [20, 21] proposes a semantic 3. The SHARE Framework similarity function that takes into account the patients medical profiles and shows its superiority over traditional In this section, we describe the SHARE framework that similarity measures in group recommendations. [22] develops our recommendation system. We begin by defocuses on the notion of fairness, devising an aggregation scribing the collaborative filtering approach that we used method for ensuring that if the group recommendation to generate recommendations, secondly, we discuss a list provides a high relevant document for a patient, then personalized recommendation method. After that, we that patient may be tolerant of the existence of documents describe the application of a personalized filtering apthat are not relevant to him/her. More recently, [23] proach. Finally, we discuss the explanations that SHARE exploits as well additional properties for producing group provides to users about why they are receiving each recrecommendations, like the education and health literacy ommendation. level, and the psycho-emotional status of the group.
As concerns, to recipe recommendations based on di- 3.1. User-Based Collaborative Filtering etary preferences or restrictions, Agapito et al. [24] proposed a personalized recipe RS that takes into account To generate recommendations for a given user, we apply users’ health profiles and chronic diseases, such as CKD, CF using the user’s ratings and the ratings of other similar hypertension, and diabetes. In a similar way, Yang et al. users. [25] developed a food RS that takes into account users’ Before applying the similarity measure, we normalize nutritional expectations, dietary restrictions, and fine- the ratings of all users by subtracting the mean rating grained food preferences. of each user from their ratings. This has the efect of
Regarding the incorporation of user-generated data, centering the ratings around the mean, with positive such as recipe ratings, and reviews into the recommen- ratings indicating ones that are higher than the mean and negative ratings indicating ratings that are lower that the user would give to each recipe. The formula for than the mean[29].
Normalizing the ratings in this way helps to take into account the diferences in the absolute rating scales used by diferent users. For example, one user may tend to rate all recipes as 5 stars, while another user may rate the same recipes as 1 star. Without normalization, the similarities between these two users would be artificially low due to the diferences in their rating scales. Normalization assists in fixing this by adjusting the ratings to a common scale. Normalization is an important step in the CF process because it helps to ensure that the similarities between users are based on their relative preferences rather than their absolute rating scales [29].
To identify similar users, we first compute the similarity between each pair of users based on the cosine similarity measure. The cosine similarity measures the similarity between two non-zero vectors of inner product space and is defined as the cosine of the angle between the vectors.1 The resulting similarity ranges from −1 meaning completely dissimilar, to 1 meaning completely similar, with 0 indicating orthogonality, while in-between values indicate intermediate similarity or dissimilarity.
We use it to measure the similarity between the ratings of two users.2 The formula for the cosine similarity is as follows: the predicted rating of a recipe, ,by user is as follows: ( , ) = + ∑ ( ∑ , ) (
, ) ( , )
(2)
In this formula, the (“neighbors”) are the users who are most similar to the target user , as determined by the cosine similarity measure. The predicted rating is a weighted average of the ratings of the neighbors, with the weights being the cosine similarities ( tween the users. Finally, the average rating target user is included in the prediction to take into account the fact that diferent users may have diferent
of the overall rating tendencies. Because as we discussed above for ”Normalization”, one user may tend to give higher ratings to recipes overall, while another user may tend to give lower ratings. Including the average rating of the user in the prediction formula helps to adjust for these diferences in rating tendencies so that the predictions are more accurate and reflective of the user’s true
, ), bepreferences. to the user.
Once the predictions have been computed for all recipes, we can rank them and recommend the top-rated recipes (, ) =
∥ ∥∥ ∥
To compute the cosine similarity, we first convert the
ratings of each user into a term-frequency representation,
which represents the frequency with which each rating
appears in an individual’s rankings3. For example, if
someone rated 3 recipes as 5 stars and then 2 more at 4
with 1 left over at 3; their corresponding term-frequency
would be [
Once we have the term-frequency representations of the ratings of two users, we compute the dot product of the vectors by multiplying the corresponding elements of the vectors and summing the results. The dot product is then divided by the product of the magnitudes of the vectors to give the cosine similarity. The magnitude of a vector is the square root of the sum of the squares of the elements of the vector 4. Intuitively, we use the cosine require the ratings to be normally distributed5. similarity because it can handle sparse data and does not
To generate recommendations for a given user, we use the ratings of a set of similar users to predict the rating
1https://medium.com/@riyasisonline/cosine-similarity-is-adefinition/
(1)
To take into account the health needs of the individual user, we incorporated personalization techniques that allow the system to consider the user’s health history.
To achieve this, we combined collaborative filtering, content-based filtering, and knowledge-based methods.
CF relies on the preferences of similar users to recommend items [30], while CB uses the characteristics of the item itself to make recommendations [31]. Knowledgebased recommendation systems use a combination of explicit knowledge about the items and the preferences of the users to make recommendations [32].
For the content-based component, tags were extracted from recipe descriptions including nutritional information and other relevant characteristics of the recipes.
These features formed a vector for each recipe, which incorporated the recipe’s nutritional information and other relevant attributes.
For the knowledge-based component, we identify the specific nutrients that are suitable for every chronic health condition supported by SHARE. The data was collected by oficial statistics 6789. We used this information to information and use it to generate human-readable explacalculate the nutritional profile that is most suitable for nations for the recommended recipes. To do this, SHARE the user’s health needs, taking into account their specific uses a database for storing and retrieving information health condition(s). We then used this nutritional pro- about recipes, users, and chronic diseases, allowing the ifle to create a target vector, which represents the types system to generate explanations that are personalized to of recipes that are suitable for the user based on their the user based on their preferences and specific health specific health needs. needs.
Finally, we used the cosine similarity between the fea- For example, if a recommendation is made to a user ture and target vectors to identify recipes that are less with obesity, the explanation might highlight the fact suited to the nutritional profile of the user and we exclude that the recommended recipe is low in saturated fat and them from the collaborative filtering we apply after. high in fiber, and that these characteristics are beneficial
Once the CF process is finished, the system will have for managing obesity. produced a list of recommended recipes that are cus- The justifications are provided by the system through tomized to meet the individual’s health requirements and tables like Table I below. personal preferences. The explainability component of our recommender system is designed to provide users with a greater under3.3. Personalized Filtering standing of the reasoning behind the recommendations and to enable them to make more informed decisions about which recipes to try.
We apply a personalized filtering technique in both methods to improve the accuracy and relevance of recommendations in the two methods we discussed above.
SHARE ofers personalized filtering, allowing users 4. Experiments to customize the recommendations based on their tastes and dietary needs. Through personalized filtering, users 4.1. Dataset / Set-Up are able to narrow down their recommendations based The dataset used for our experiments was obtained from on tags extracted from recipe descriptions (e.g. vegan or food.com10 containing real-world recipe ratings collected gluten-free) and desired nutritional values like calories, from numerous users, forming an ideal source for our RS. protein, and saturated fat. The dataset consisted of user IDs and associated recipe
This elevated level of customization helps ensure that IDs with corresponding rating scores on a scale of 1 to 5 our recommended recipes accurately meet individual that reflected the level of satisfaction expressed by each preferences, resulting in highly relevant and useful re- individual user towards the respective recipe in question. sults for maximized satisfaction. All entries were arranged into a CSV file wherein the ifrst column corresponded to the ID signifying specific 3.4. Explainability individuals, followed by a column representing the recipe identifiers, and the last one corresponding to the given One of the challenges of using machine learning methods rating. for recommendation systems is the lack of explainability By utilizing the health history of each user, we are able of the results. This can make it dificult for users to to make more personalized recommendations tailored understand why a particular recommendation was made. to their health needs. To do so, we gather data from
To address these issues, SHARE includes an explain- oficial statistics 111213141516 and we assigned a collection ability component that provides users with a clear and of Boolean attributes outlining whether the users of the concise explanation of the reasons behind each recom- dataset have any chronic conditions or not. This allows mendation. us to generate suggestions that take into account people’s
The explanations are generated using a combination medical history making it easier for them to manage of natural language and domain-specific knowledge that includes recipe tags, nutritional properties as well as information about the preferences and dietary restrictions of the users. The system is designed to process this 6https://www.greenfacts.org/en/diet-nutrition/index.html/ 7https://www.hopkinsmedicine.org/health/conditions-anddiseases/cancer/cancer-diet-foods-to-add-and-avoid-during-cancertreatment/
8https://stanfordhealthcare.org/medical-clinics/cancernutrition-services.html/ 9https://www.eatingwell.com/our-food-nutrition-philosophy/ 10https://www.kaggle.com/datasets/shuyangli94/food-comrecipes-and-user-interactions
11https://www.who.int/news/item/04-03-2022-world-obesityday-2022- accelerating-action-to-stop-obesity 12https://diabetesresearch.org/diabetes-statistics/ 13https://www.cdc.gov/heartdisease/facts.htm 14https://www.wcrf.org/cancer-trends/worldwide-cancer-data/ 15https://www.who.int/news-room/fact-sheets/detail/oralhealth
16https://www.bonehealthandosteoporosis.org/wp- content/uploads/2015/12/Osteoporosis-Fast-Facts.pdf • The first question assessed the accuracy of the results in each method separately. This question asked participants to rate on a 5-point rating scale, The questions presented to the real users are shown in their satisfaction with the provided recommen- Table III. dations based on users’ past behavior. This can provide valuable insights into how well the RS is 4.3. Results meeting the needs of the users. • The second question assessed the personalization The results of the survey presented in the previous secof the results in each method separately. This tion are shown in the sequel. chronic illnesses when proceeding with any decisions relating to food.
In the experiment, four methods were utilized including, Collaborative Filtering(CF) discussed in subsection III/A, Health Collaborative Filtering(HCF) discussed in subsection III/B, Personalized Collaborative Filtering(PCF) which combines the CF method and Personalized filtering discussed in subsection III/C, Health Personalized Collaborative Filtering(HPCF) which combines the HCF method and Personalized filtering discussed in subsection III/C.
The components of each method are shown in Table
To evaluate the usability of our four methods, we conducted a survey with 40 real users with 40% being between 18-24 years old, 40% between 25-49 years old, and 20% being 50-59 years old. The participants had a diverse range of educational and professional backgrounds, including 40% with a computer science background, 30% with a healthcare background, 10% with a cooking background, and 20% with a general background.
The survey consisted of 6 questions, which were designed to assess diferent aspects of the methods’ usability. Each question focuses on a specific aspect of usability. Specifically: question asked participants to rate on a 5-point rating scale whether the recommendations provided were helpful and relevant for the specific health problem they are facing. This question can help us understand whether the system is efective at providing useful and relevant recommendations to users who are seeking recipes suitable to specific health problems. • The third question assessed the user acceptance of the results in each method separately. This question asked participants to state how many of the suggested recipes they find appealing. This question can help us understand whether the system is efective at providing recommendations that the users are interested in or that meet their needs. • The fourth question asked participants to state which method they preferred. This question can help us understand which method is the most famous among users. • The fith question assessed the overall coverage of the results. This question asked participants to rate on a 5-point rating scale, their satisfaction with the variety of provided recommendations. By assessing the coverage of the results, we can get an indication of how well the recommender is able to suggest a wide variety of recipes to users. • The sixth question assessed the overall explainability of the system. This question asked participants to rate on a 5-point rating scale the justification provided for the recommended recipes is sufifcient for them to make a decision and whether the popularity of the recipes influences their decisionmaking process. These questions can help us understand how users make decisions based on the recommendations provided, and whether the system is providing suficient information to support those decisions.
Question
The results of the first question, focusing on the accuracy of the results, are presented in Fig. 1. As shown, methods HCF and HPCF have the most accurate results with HPCF results being slightly more accurate than the results of HCF. The PCF method has a moderate rating compared to the above. On the other hand, the CF method has the lowest rating among all.
g n it a r e g a r e v A 4.4 4.2 4 3.8
CF
PCF
HCF
HPCF
The results for the second question, focusing on the personalization of the results, are shown in Fig. 2. According to the figure HCF and HPCF have more personalized results with HPCF results being slightly more personalized than the results of HCF. The PCF method has a moderate rating compared to the above. On the other hand, the CF method has the lowest rating among all.
The results of the third question, which asked about user acceptance of the results, as shown in Fig. 3, indicate that HPCF has the most appealing results. The HCF method has a moderate rating compared to HPCF. Finally, the other two methods have a considerably lower score than the above methods.
The results of the fourth question, which asked about the most favored method, as shown in Fig. 4, indicate that HPCF is clearly the most famous method among users. The other three methods have much lower results, g n it a r e g a r e v A g n it a r e g a r e v A 5 4.5 4 4.2 4 3.8 3.6 3.4 3.2
CF
PCF
HCF
HPCF with CF having an overall score equal to zero and PCF also close to it.
Finally, the results of questions 5 and 6, which asked about the overall coverage and justification of the results of all methods, are shown in Table IV. The results indicate that the system provides a wide variety of recipes to users and excellent justifications.
Overall, the results of the survey indicate that the
CF
PCF
HCF
HPCF 5. Conclusion and Future Work In this paper, we presented a personalized recommendation system that recommends recipes to users based on their health history and the preferences of similar users. Overall, the SHARE framework combines user tastes and nutritional information about the recipes in order to provide recommendations for recipes that meet the user’s preferences and specific health needs. We also ofer personalized filtering for the users of the system. Finally, we evaluate its usability through a series of experiments on a large real-world data set of recipes. Our experiments demonstrate the system’s ability to provide highly relevant personalized recommendations.
There are several directions in which the work presented in this paper could be extended in the future. One possible extension is to incorporate additional types of user data, such as age, gender, allergies, exercise habits, and physical activity levels, to make more informed recommendations. Another possible future direction is enriching the RS by including more factors, such as cultural background or social connections. This could allow the system to suggest recipes that are more likely to be wellreceived by the user’s friends and family. Finally, we believe it is worth trying to expand the system by considering other factors beyond the user’s health history, such as the user’s location, the season, the availability, and the cost of ingredients, to make more contextually relevant recommendations. Overall, there are many exciting possibilities for improving the performance and usability of the recipe recommender presented in this paper.