Intelligent Systems with Applications 16 (2022) 200157

Contents lists available at ScienceDirect

Intelligent Systems with Applications

journal homepage: www.journals.elsevier.com/intelligent-systems-with-applications

An effective explainable food recommendation using deep image clustering

and community detection
Mehrdad Rostami a,∗ , Usman Muhammad a , Saman Forouzandeh b , Kamal Berahmand c ,
Vahid Farrahi a,d , Mourad Oussalah a,d
a
Center for Machine Vision and Signal Analysis (CMVS), University of Oulu, Finland
b
School of Mathematics and Statistics, University of New South Wales, Sydney, NSW, Australia
c
School of Computer Sciences, Science and Engineering Faculty, Queensland University of Technology (QUT), Australia
d
Research Unit of Medical Imaging, Physics and Technology, University of Oulu, Finland

A R T I C L E I N F O A B S T R A C T

Keywords: In food diet communication domain, images convey important information to capture users’ attention beyond the
Recommender system traditional ingredient content, making it crucial to influence user-decision about the relevancy of a given diet. By
Food recommendation using a deep learning-based image clustering method, this paper proposes an Explainable Food Recommendation
Explainable artificial intelligence
system that uses the visual content of food to justify their recommendations. n the recommendation system.
Deep learning
Especially, a new similarity score based on a tendency measure that quantifies the extent to which user
community prefers a given food category is introduced and incorporated in the recommendation. Finally, a
rule-based explainability is introduced to enhance transparency and interpretability of the recommendation
outcome. Our experiments on a crawled dataset showed that the proposed method enhances recommendation
quality in terms of precision, recall, F1, and Normalized Discounted Cumulative Gain (NDCG) by 7.35%, 6.70%,
7.32% and 14.38%, respectively, when compared to other existing methodologies for food recommendation.
Besides ablation study is performed to demonstrate the technical soundness of the various components of our
recommendation system.

1. Introduction of individuals, this strategy seeks to provide them with a range of appro-
priate food products to enable them to perform their everyday activities
Throughout human history, food has been a significant aspect of life. and beyond while meeting their dietary preferences (Ali et al., 2018,
For human survival, humans had to recognize and hunt for food at first. Chen & Toumazou, 2019, Vaishali & Shukla, 2019).
Nowadays, diet choice is already becoming more and more important During the last decade, worldwide web technology and cellphone
in order to provide basic nutrition, calories, taste, mental well-being, equipment have grown exponentially (Halim et al., 2019, Halim &
and sociocultural factors (Kim & Chung, 2020, Premasundari & Yamini,
Rehan, 2020). People nowadays have access to massive amounts of
2019, Srilakshmi et al., 2022, Tran et al., 2021). The rise in obesity
interactive digital food information from a variety of sources, includ-
and diabetes is largely caused by the way we eat (Bishop et al., 2021,
ing blogs, social networks, and consumer product reviews (Ding et al.,
Molina-Ayala et al., 2022, Zhu et al., 2022). The Global Burden of
2022, Kang et al., 2022, Samad et al., 2022, Zhang et al., 2022). As this
Disease Study stated that dietetic patterns contribute significantly to
malnutrition, fatness, and adiposity thresholds and those unhealthy di- expansion offers users more choices, it also makes it more challenging
ets cause 11 million preventable early deaths every year (Wang et al., for them to choose from a wide range of food options. Due to this, food
2019). As a new branch of science, food recommendation systems are recommendations have become progressively important for meeting po-
beginning to emerge to address these issues (Agapito et al., 2017, James tential customer requirements and, as a result, for assisting consumers
et al., 2018, Lee et al., 2020, Norouzi et al., 2017, Subramaniyaswamy to find appropriate food recommendations easily (Gusnedi et al., 2022,
et al., 2019). In order to meet the basic biological/physiological needs Subramaniyaswamy et al., 2019).

* Corresponding author.
E-mail address: Mehrdad.Rostami@oulu.fi (M. Rostami).

https://doi.org/10.1016/j.iswa.2022.200157
Received 16 September 2022; Received in revised form 31 October 2022; Accepted 21 November 2022
Available online 23 November 2022
2667-3053/© 2022 The Authors. Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

available relevant food images to perform a tailored clustering task,

which is incorporated to derive the food recommendation.
• eXplainable Food Recommendation (XFR): Using rule-mining
based approach, an effective explainable module is incorporated in
our food recommendation system, which explains the reasons for
recommending a specific recommendation to the target user. Espe-
cially, the explainable module identifies rules that predict whether
a user will likely be interested in food that he/she tasted (or rated)
previously. To the best of our knowledge, this is the first work that
incorporates explainability into a food recommender system.
• Time-aware Food Recommendation (TFR): A novel time-aware
food recommendation that considers the temporal information of
ratings in food rating prediction is introduced. In our developed
system, different from previous food recommender systems (Gao et
al., 2019, 2022, Princy et al., 2021, Teng et al., 2012), a weight-
ing mechanism is designed to handle the importance of the time
factor for different ratings where old ratings are assigned lower
importance scores than newer ones. This time-aware function in-
corporates users’ preferences into the recommendation process dy-
namically.
• User Communities and Food Groups-aware Food Recommen-
dation (UCFG): Acknowledging the importance of community as-
pect in influencing the user’s diet behavior, our model takes into
account communities generated by clustering users’ taste prefer-
Fig. 1. A simple example of a visual illustration of food ingredients correspond-
ences. For this purpose, a new measure is introduced to calculate
ing to a user who likes chicken wing. the tendency of individuals in each user community to each food
group that can be used when recommending new foods to the users.
1.1. Gap analysis
1.2. Research questions
There have been several food recommendation systems developed
in recent years to predict and/or guide people’s choices based on some In the current paper, the following research questions are addressed
predetermined set of criteria. Despite the fact that previous food rec- to mitigate previous shortcomings of food recommendation models:
ommender systems achieved good performance in terms of learning
user’s preferences by mapping historical interactions with food items • RQ1: How to integrate food images in the food-recommendation
and recipes, these systems still suffer from two significant limitations. pipeline? To answer this research question a new deep learning
The first one is related to ignoring the potential impact of food im- model that clusters food images are integrated with the time-aware
ages, if any, in the food recommendation process. Indeed, food images user similarity (to account or user’s rating) as well as user’s com-
can convey more attractiveness than just ingredient list due to their po- munity is suggested in Section 3.
tential to influence consumer’s feeling and behavior. For instance, it is
• RQ2: How can the recommendation of the developed recommender
possible to taste some ingredient quite differently depending on how
system be justified to user? To answer this research question, a
they are cut and cooked, while the food content is still unchanged. This
rule-based approach using associative rule-mining method has been
testifies of the potential of food layout that can be illustrated through
suggested in Section 3.5.
its visual content. Therefore, as phrased in Chinese proverb “a picture
• RQ3: What is the impact of employing user communities and food
is worth a thousand words”, it is intuitively much easier for a user to
groups on the final performance of the recommender system? To
determine the taste of a recipe from its visual content than from its tex-
answer this research question, a set of experiments have been
tual description. As a result, it is vital that the visual semantics of food
conducted where results with and without visual components are
images, whenever available, are properly taken into account when pro-
recorded. This is reported in Section 4 of this paper.
viding food recommendations. An example is shown in Fig. 1, which
displays a user and his four recent food choices. It is evident from his
tasted recipes that this user likes foods that include chicken wings. On The rest of this paper is organized as follows. We review the lit-
the other hand, the lack of explanation or transparency in the previ- erature in Section 2. The developed system is detailed in Section 3.
ous food recommender systems is another major limitation/challenge Experimental results and their discussions are provided in Section 4.
that restricted the widespread of food recommender systems. Indeed, Finally, Section 5 concludes the paper and provides some future direc-
the lack of compelling explanations for the generated food recommen- tions.
dation may render the user unwilling to adopt it or follow its update. As
a result, a good and efficient food recommendation is one that explains 2. Literature review
to each individual why a specific food diet is recommended to him/her.
We have addressed the above-mentioned limitations in this study by
developing a novel explainable visual food recommender system with This section first provides a brief introduction to recommender sys-
the following unique characteristics: tems and then distinguishes the state-of-the-art in Explainable, Image-
aware, Time-aware, and User/Food group aware recommendation sys-
• Image-aware Food Recommendation (IFR): In contrast to previ- tems. Each of these groups is further discussed in its own subsection.
ous food recommendation systems (Chavan et al., 2021, Rostami Finally, the studied food recommendation systems and their methods,
et al., 2022, Trattner & Elsweiler, 2017) that neglect food images, motivations and limitations are summarized in Table 1, while providing
we developed a novel convolution neural network model that uses insights into the contribution and limitation of our proposed model.

2
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

2.1. Recommender systems performance. Moreover, in this study, a visual representation of the
recommendations provided in the proposed framework enables direct
Recommender systems are among active research fields in the last interpretation. Additionally, explanations may aid in improving user ex-
few decades in academia, business and service industry (Karimi et al., perience and detecting system defects. By improving the transparency
2018, Kunaver & Požrl, 2017). From a methodological approach per- of the representation learning process, the authors of Liu et al. (2020)
spective, we can distinguish three main streams of techniques employed developed an explainable recommendation model. The concept of graph
in recommender systems: Content-Based Filtering (CB), Collaborative convolution is revised to provide discrimination between information
Filtering (CF), and Hybrid Filtering (HF). from different layers in order to overcome traditional models’ represen-
Content-based filtering recommendation system (Albatayneh et al., tation entanglement problems.
2022, Van Dat et al., 2022) aims to recommend items to users by tak-
ing into account the profile of user’s preference and content of items.
It constructs a user’s preference based on product information that the 2.3. Image-aware recommender systems
user has previously liked, calculates the similarity between the candi-
date item and the preference model, and recommends the item that is Over the last few years, image-based recommendation models have
most similar to user’s preference. In essence, CB system targets items become increasingly popular (Hiriyannaiah et al., 2020, Sulthana et al.,
that the user previously liked. 2020, Ullah et al., 2019, Yu et al., 2021). Considering the image of
Collaborative Filtering (CF) technique recommends items accord- the item is what first catches the eye, it seems intuitive to consider the
ing to previous user interests with similar flavors to the target user pictorial representation as the gold standard for analyzing the item’s
(Forouzandeh et al., 2022), regardless of the content of these items. content. The visual aspect of information plays an important role in
Especially, in CF, the historical relationships between users and items human decision-making. This aesthetic factor is very important to pre-
are mined to determine the preferences of these users, then generate dict user preferences. Based on the image of the item, these models
recommendation by ranking candidate items according to predicted are able to determine the content of that item (Yu et al., 2021). Tradi-
preferences. CF provides assistance in making decisions based on the tionally, image-based recommendation system uses the images of users’
opinions of other users. We distinguish user-based CF and item-based rated items to determine the most similar users. Then it returns the
CF. User-based CF represents each user as a vector of ratings, and pre- most likely items to the target users based on the images liked by their
dicts the user’s missing rating on a new item based on the weighted neighbors.
average of other users’ rating on the same item. Item-based CF repre- In Yu et al. (2021) the use of aesthetic features in recommender sys-
sents each item as a vector of ratings, and then predicts the missing
tems using implicit feedback datasets was investigated. In order to cap-
rating according to the weighted average of ratings from similar items.
ture the aesthetic preferences of users, the authors developed a method
Other important strategies to create harmony are covered by hy-
that incorporated aesthetic features into a tensor factorization model
brid methods (Forouzandeh, Berahmand, Nasiri, et al., 2021), which
and leveraged visual information to optimize it. Hiriyannaiah et al.
combine both collaborative filtering and content-based approaches by
(2020) developed an image-aware recommender system using the Deep
complementing the weaknesses of one method by another one. This is
Visual Ensemble similarity metric and Convolutional Autoencoder. In
referred to hybrid filtering technique, which expects to yield more accu-
their study, the deep learning-based similarity measure is utilized to
rate and practical recommendations than a single algorithm (Forouzan-
find similarities between trained feature vectors and the target feature
deh, Berahmand, Rostami, 2021, Forouzandeh, Rostami, et al., 2021).
vector. Using multi-view visual information and implicit feedback data,
the authors of Zhang, Luo, et al. (2020) developed a new factoriza-
2.2. Explainable recommender systems
tion model for recommending restaurants. In their study, using a deep
Explainable recommender systems aim to propose approaches that convolution network, visual features (visual information) are extracted
develop not only accurate recommendation but also appropriate expla- from images and integrated into collaborative filtering-based recom-
nations, such as why specific items are recommended to a given target mender system.
user, which, ultimately, yields more transparent, persuasive, effective,
trustworthy, and user-friendly result. 2.4. Time-aware recommender systems
In the literature of explainable AI, we distinguish model-intrinsic (or
pre-hoc) and model-agnostic explainability (Lipton, 2018). The former
In time-aware recommendation systems, users’ preferences are mod-
advocates an AI system that explains its decisions based on its infer-
eled by taking into account temporal dynamics. Temporal models em-
ence systems (Zhang et al., 2014), while model-agnostic explainabil-
ity focuses on explaining model outputs without knowing the internal phasize how users’ preferences change over time and context. The drift
mechanism of the model (Peake & Wang, 2018). Especially, the model- concept has been addressed by developing different time-aware recom-
agnostic methods allow the decision mechanism to be a black box. mender systems over the last few years (Ahmadian et al., 2022, Cui et
These two types of explanations are profoundly rooted in human cogni- al., 2020, Noulapeu Ngaffo et al., 2021, Rostami et al., 2022, Sánchez-
tive psychology – decisions can sometimes be based on careful, rational Moreno et al., 2020).
reasoning, where we do explain why we do so; while, in some other The authors of Sánchez-Moreno et al. (2020) developed a time-
cases, we make decisions first, and then seek explanations for them to aware music recommendation system based on modeling implicit user
justify or support them further (Zhang & Chen, 2020). preferences over time. To provide users with better recommendations,
In terms of methodological framework, explainable recommenda- they used the collaborative filtering method to capture users’ listening
tion models can be divided into seven categories: Factorization Models, habits on a daily basis. Two unified methods are used by Kefalas and
Topic Modeling, Graph-based Models, Deep Learning-based Models, Manolopoulos (2017) to develop a time-aware recommendation system
Knowledge Graph-based Models, Rule Mining Models and Post Hoc that considers the spatial, textual, and temporal elements simultane-
Explainable Recommendations (Zhang & Chen, 2020). In order to ex- ously. Their method also evaluated the impact of time on different
plain recommendation models, the users’ or items’ content informa- intervals of time by taking into account the temporal dimension. In
tion is used to facilitate their decision-making process. An improved Zhao et al. (2021), the user bias is considered to capture the preference
knowledge graph attention network model is used in Shimizu et al. change of users. In the context-aware recommendation system, the au-
(2022) to develop a novel explainable recommendation model that uti- thors first examined the time-varying impact on user bias and item bias
lizes the content information of items to achieve high recommendation and then developed a time-varying bias tensor factorization.

3
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

2.5. User/item group-aware recommender systems In Maia and Ferreira (2018), a context-aware food recommenda-
tion system based on a matrix factorization and feature engineering
Community detection algorithms were used in recommendation sys- was developed for well-being care applications, using mobile devices
tems to group either similar users or similar items (Deebak & Al- and medical records. Food recommendations from nearby food places
Turjman, 2020, Rostami et al., 2022, Viktoratos et al., 2018, Wang et were made available to users, and they can order healthy foods for the
al., 2020). Community detection techniques are shown to increase the table in real time based on available suggestions.
accuracy of rating prediction in these systems and improve the han- In Thongsri et al. (2022), the authors developed a personalized
dling of cold starts and sparse data issues. Chen and Toumazou (2019) healthy food recommendation system based on a combination of collab-
suggested personalized recommendation system that is achieved by gen- orative filtering and the knapsack method. The results of their research
erating a time-aware matrix to better capture user’s interest. Authors indicate a full user satisfaction and screen design efficiency.
of Viktoratos et al. (2018) suggested a context-aware recommendation As the main objective of our research is to develop a novel food
system by combining community detection and association rule mining recommendation system using deep learning-based clustering of food
techniques. Graph embedding-based methods are developed in Zhang, images, we will thereby review some of the previous deep learning-
Qu, et al. (2020) for detecting group shilling attacks in collaborative based recommender systems as well. In Naik (2020), the authors pro-
filtering recommendation systems. In their paper, the authors first an- posed a deep learning-based trained recommendation system based on
alyzed user ratings to create a user network, then they embedded each the recommendations received by the customer who has already used
node in the user relationship graph with a low-dimensional vector rep- the product. In their approach, each person has his own eating pat-
resentation. Then they used a clustering method to identify candidate terns based on the selections and disapproval of the user, indicating
communities based on the characteristics of the produced users. that a personalized diet is essential. The developed food recommenda-
tion method uses a deep learning algorithm and a genetic algorithm
2.6. Food recommender systems to provide the best possible recommendation. Using Graph Convolu-
tional Network, in Gao et al. (2022), a new food recommender system
Since their introduction and popularity in entertainment industry, based on the ingredient-ingredient, ingredient-recipe, and recipe-user
recommender systems have gained increasing attention in health ser- relationships was developed. To describe high-order connectivity across
vices, especially, with their potential to guide users towards healthy different food-related relationships and enhance representations, the
lifestyle. Besides, due the substantial increase in lifestyle-related dis- authors employed information propagation mechanisms and adopted
eases, such as diabetes and obesity, which can cause several regular multiple embedding propagation layers. In Gao et al. (2019), a ded-
disorders, the problem of choosing appropriate diet is of paramount icated deep learning-based Hierarchical Attention model for the food
importance. This sheds light on the importance of food recommender recommender system was developed. Their system was able to: 1) cap-
systems that can guide user towards more healthy diet while accommo- ture the collaborative filtering effect like what similar users eat; 2) infer
dating to some extent his/her food preferences. The majority of current a person’s preference based on the ingredients in their food; and 3)
studies in the food domain focused on providing recommendations for determine the type of food the users prefer. In Aditya et al. (2021),
favorite food items to users according to their preferences (Lee et al., a real-time food recommendation system utilizing a machine learning
2020). model was suggested to present the dishes to users based on their past
The study of Min et al. (2019) reviewed previous food recommender orders. The solution has been implemented via a mobile application
systems as well as their theoretical frameworks, existing solutions, and made using Flutter app.
challenges. The authors stressed on the multimodal aspect of food rec- According to the gap analysis extracted from the literature on food
ommender system when considering the variety of contextual informa- recommender systems, the area of food recommendation presents am-
tion that can infer taste, ingredients and food content as well as the ple opportunities for scientific and research communities to explore its
variety of approaches that can be employed in user’s profile descrip- various directions. Table 1 summarizes the studied food recommender
tion. Starke and Trattner (2021) suggested how healthy food choices systems and their method, contributions and limitations. Based on four
can be supported by integrating new approaches into the presented con- distinguished factors consisting of Explainability, Image-aware, Time-
tent and decision context. The authors gave a multi-list recommender aware, User communities and Food groups-aware, our proposed food
interface to help healthy food generation based on Netflix user inter- recommendation system is shown to be the only one that considers all
face. In Ge et al. (2015), a novel system that allows users to hover the four aforementioned aspects simultaneously. It is worth mentioning
over their preferences and health was developed. The described sys- that none of the above-mentioned food recommendation models con-
tem has been developed on the Android platform. Their work presented sidered explainability aspect in their recommendation procedure. This
a personalized health-aware food recommendation system named Mar- is mainly due to the fact that these models focused primarily on im-
ket2Dish. The primary purpose of this system is to help people find proving the accuracy of the recommendations by accounting only for
out personalized food and maintain a healthy diet, thus avoiding the factors that directly affect accuracy. However, in the food recommen-
disease caused by unhealthy eating patterns. Authors of Sookrah et dation context, explainability is often critical as it yields more trust
al. (2019) presented a DASH diet recommendation system to suggest recommendations. Therefore, in this paper, we develop an effective
healthy menus and dishes. The recommended dishes aim to assist a explainable food recommender system that takes into account simul-
hypertensive person in controlling his/her diet and prevent him from taneously both user ratings and food images.
getting health complications. In Oh et al. (2010), a context-aware food
recommendation system was proposed for well-being care applications. 3. Developed system
Their proposed approach, called u-BabSang, provides individualized
food recommendation lists at the dining table using dietary advice of Traditionally, collaborative filtering has failed to capture visual data
a typical Korean medical text. associated with items where recommendation models are expected to
Rehman et al. (2017) emphasized that selecting a diet that is suit- perform better when visual features of items are used (Hiriyannaiah
able for patients must meet their nutritional needs. They developed a et al., 2020). Our model incorporates both food image content-based
cloud-based food recommendation system, called Diet-Right, for dietary model and community-aware recommendation, as will be detailed in
recommendations based on users’ pathological information to manage this section. Besides, due to personalized nature of diet, eating habits
this issue. The model used an ant colony algorithm to generate an opti- and social community aspect where each group of people tends to taste
mal food list and recommended suitable foods according to the values a certain type of food only, our model considers food grouping and
of pathological reports. user community detection simultaneously. The developed model called

4
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

Table 1
The summarized descriptions of the studied food recommender systems.

Reference Method Contributions Limitations

Starke and Trattner (2021) Multi-list recommendation Food similarity calculated using multiple attributes It is not explainable. Food images are also ignored
(ingredients, nutrients, etc)
Ge et al. (2015) Matrix factorization Health factors of foods taken into account It is not explainable and Time-aware. Food images and
food groups are also ignored
Sookrah et al. (2019) Content-based Recommend healthy menus and dishes. It is not explainable and Time-aware
Oh et al. (2010) Context-aware Considering user profile, physiological signals, and Time and food images are not considered
sensed environmental data.
Rehman et al. (2017) Ant colony optimization-based Recommending foods that meet the nutritional needs It is not explainable. Time and food images are not
of patients considered
Maia and Ferreira (2018) Context-aware Develop a novel matrix factorization and feature Time and food images are not considered
engineering-based model
Thongsri et al. (2022) Collaborative filtering Recommend health food using knapsack-based Food content is not considered. Transparency and
algorithm explanation are lacking
Naik (2020) Collaborative filtering Recommend favorite foods using deep learning and Lacking of explainability. Not considering the food
genetic algorithms content
Gao et al. (2022) Hybrid model Exploiting ingredient-ingredient, ingredient-food, and Due to lack of consideration of the time factor, user’s
food-user connection dietary changes cannot effectively be handled
Gao et al. (2019) Hybrid model Hierarchical attention graph for image-aware Lacking of explainability. Not considering the time
recommendation factors of user rates
Aditya et al. (2021) Collaborative filtering Developing a deep matrix factorization model Time and food images are not considered. It is not
explainable
Developed system Hybrid model Explainable and time-aware model that use recipes Our developed system cannot guarantee that the
images to identify food groups recommended foods are healthy

Explainable Food Recommendation by Deep Image Clustering, in short Inputs:

{ }
EFRDIC, is grouped as a hybrid recommender system that utilizes the - User set: 𝑈 = 𝑢1 , 𝑢2 , 𝑢3 , … , 𝑢𝑁
{ }
advantages of both collaborative filtering and content-based models. - Food set: 𝐹 = 𝑓1 , 𝑓2 , 𝑓3 , ..., 𝑓𝑀
{ }
EFRDIC recommends to each user a set of foods that he/she may be - Food image set: 𝐹 𝐼 = 𝑓 𝑖1 , 𝑓 𝑖2 , 𝑓 𝑖3 , ..., 𝑓 𝑖𝑀
most interested in. - User-Food rating matrix: 𝑅 = {1, 2, 3, 4, 5, −}𝑁×𝑀
The overall flow diagram of the developed system and four distin- - Rate given to food 𝑓𝑖 by user u: 𝑟𝑢,𝑖
guished factors of Image-aware Food Recommendation (IFR), eXplain-
Output:
able Food Recommendation (XFR), Time-aware Food Recommendation { }
- User communities: 𝑈 𝐶 = 𝑢𝑐1 , 𝑢𝑐2 , 𝑢𝑐3 , … , 𝑢𝑐𝐾
(TFR) and User Communities and Food Groups-aware Food Recommen-
- Community of users that user 𝑢 belongs to: 𝐶𝑜𝑚(𝑢)
dation (UCFG) are shown in Fig. 2. More precisely, EFRDIC has two { }
- Food groups: 𝐹 𝐺 = 𝑓 𝑔1 , 𝑓 𝑔2 , 𝑓 𝑔3 , … , 𝑓 𝑔𝐶
simultaneous steps: (1) Deep learning-based food image clustering, and
- Groups of foods that food 𝑓𝑖 belongs to: 𝐺𝑟𝑜𝑢𝑝(𝑓𝑖 )
(2) Time-aware user community detection, followed by two other steps: ( )
- Rating Prediction Function: 𝑝̂𝑖 𝑢𝑗 = 𝑓 (𝑈 , 𝐹 , 𝐼, 𝑅)
(3) User community-food group tendency estimation, (4) Explainability { }
- Recommendation: 𝐹̂ = 𝑓̂1 , 𝑓̂2 , 𝑓̂3 , ..., 𝑓̂𝐿
module using associative rule mining. In the first step, keeping in mind
- Explanation: 𝑓𝑖 → 𝑓𝑗 refers to the rule that states that if a user tasted
that “A picture is worth a thousand words”, the image of each food is
(or rated) food 𝑓𝑖 previously, he/she will also likely be interested in
used to group the food into different clusters. In the second step, first,
food 𝑓𝑗 .
the user-user similarity matrix is calculated using the recorded user rat-
ings. Next, users are represented as a weighted social network, where
3.2. Deep learning-based food image clustering
each user corresponds to a node while user-to-user similarity scores de-
note the edge weights in this newly constructed social network. Then, a
While significant attempts have been made in developing various
novel community detection method is developed for user grouping. Af-
methods for food recommendation, most of them depend on low-level
ter these two steps, a novel measure is introduced that quantifies the
features. This work proposes a robust food recommendation method
tendency of each (user) community to a given food group. Finally, uti-
which is based on visual words from deep convolutional features. Dif-
lizing this tendency function, final favorite foods are recommended to ferent from the traditional food recommendation methods in which
user (s). While the explainability module using the associative rule min- visual words are usually devised from bag-of-word model, the pro-
ing technique acts as a model-agnostic like approach. posed method introduces a deep clustering-based visual words directly
In the remainder of this section, the different phases of our food from deep convolutional features in order to provide quality recom-
recommender system are detailed. mendation services. Specifically, it is composed of three main steps:
off-the-shelf CNN feature extraction, codebook generation, and feature
3.1. Problem definition encoding. We explain below these three steps in detail.

In our food recommender system, the User set and Food set are 3.2.1. Off-the-shelf CNN features
{ } { }
indicated by 𝑈 = 𝑢1 , 𝑢2 , 𝑢3 , ..., 𝑢𝑁 and 𝐹 = 𝑓1 , 𝑓2 , 𝑓3 , ..., 𝑓𝑀 , respec- Patch sampling or feature learning is a key step for building intel-
tively, where, 𝑁 and 𝑀 stand for the number of users and foods ligent system in which edge detection, corner detection or threshold
in the dataset, respectively. Let 𝑅 correspond to user rating matrix segmentation can be utilized to extract the discriminative features. Most
that specifies the registered votes of each user for each food. Let clustering-based methods utilize different local descriptors (SIFT, SURF,
{ }
𝐹 𝐼 = 𝑓 𝑖1 , 𝑓 𝑖2 , 𝑓 𝑖3 , ..., 𝑓 𝑖𝑀 be the set of images where 𝑓 𝑖𝑘 corresponds BRIEF, etc.) in a bag-of-word framework. This is because it is still not
to the image associated with food 𝑓𝑘 in the food set 𝐹 . clear which descriptor is most suitable, and even no optimal patch size
The main function of the developed food recommender system is to can be set to choose which features are important in each given food im-
predict the rating of User 𝑢𝑗 for food 𝑓𝑖 . In overall, the input / output age. Inspired by the concept of end-to-end learning (Cheng et al., 2017),
notations of our model are summarized below: we propose to use deep learning-based clustered features, obtained from

5
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

Fig. 2. Conceptual framework of the developed model.

a pretrained CNN (Krizhevsky et al., 2012) model and then clustered 3.2.2. Codebook generation
with k-means clustering to summarize the most descriptive and salient Codebook generation is an important step in our work that takes
features throughout an image. There are two main factors behind this the convolutional features extracted from all training images and then
considered as a bag of features. This is achieved by using unsupervised
motivation. Firstly, it automatically generates feature representations
k-means clustering by grouping the data points into distinct subgroups.
from raw image pixels directly without any human supervision, which
Here, the number of clusters (k) is the size of codebook.
makes these features exhibiting more semantic properties. Secondly, the
features are closed to densely sampled SURF or SIFT features which 3.2.3. Feature encoding
make them suitable for visual word generation and the subsequent fea- This step encodes the feature descriptors into a global histogram
ture encoding. representation for each image. Specifically, feature vectors are quan-

6
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

Fig. 3. Illustration of the proposed deep learning-based food image clustering.

tized into visual words to formulate a dictionary where the image is where, 𝑡 (𝑢, 𝑖) indicates the time period of recorded rate of user 𝑢 to food
represented by the histogram of the codewords (Halim et al., 2022). 𝑓𝑖 , 𝑇 𝑃 denotes the maximum Time Period, and 𝜆 specifies a user con-
Therefore, we obtain a k-dimensional clustering-based representation trol parameter that adjusts the impact of time factor. A high (resp. low)
for each food image. Fig. 3 illustrates the overall framework of the pro- value of 𝜆 denotes a greater (resp. smaller) impact of time factor in cal-
posed deep learning-based food image clustering. It should be noted culating similarity values. In this study, the optimum configuration of
that in our model the FC6 Feature Layer is used to extract features. 𝜆 parameter has been chosen through the classical trial-and-error pa-
rameter optimization technique. The results evaluate EFRDIC model in
3.3. Graph-based overlapping user clustering terms of Precision@10, Recall@10, F1@10, and NDCG@10 for vari-
ous 𝜆 in the range of [0.5, 4] showed that in most cases, setting the
Users in many real-world social networks may belong to more than time weight parameter to 2 improved the efficiency of the food recom-
one community where various communities may overlap. Using over- mendation model. This parameter sensitivity analysis is reported in the
lapping community detection methods, variations in user interest can experimental result section.
be identified more effectively, which offers opportunities to enhance The SLPP algorithm uses listener (consumer) and speaker (provider)
performance of the food recommender-system. This also contributes to types of user. These two roles are switched based on the role of the
tackle cold-start issues. A cold-start user (or food) problem occurs when corresponding node. Overall, this overlapping community detection al-
users (or foods) have not yet accumulated sufficient rating history, gorithm includes the following steps:
which can lead to unreliable similarity scores (required for collabora-
tive filtering). • The memory of each node is initialized by its node’s id (i.e., a
The overlapping community detection algorithm employed in this unique label).
study is an extension of the Label Propagation Algorithm (LPA) named • These steps are iterated until the stop condition is satisfied:
Speaker-listener Label Propagation Algorithm (SLPA). Every node in – One node is selected as a listener.
LPA has a single label whose status is updated according to the majority – Each neighbor of the selected node sends out a single label using
label in the neighborhood. Non-overlapping communities are identified the speaking rule. Using this rule, a random label from the mem-
when the LPA algorithm converges. The idea of allowing each node to ory is selected with probability proportional to its occurrence
have multiple labels was suggested in Xie et al. (2011) to detect over- frequency.
lapping communities. For this purpose, it is crucial to determine 1) how – The listener accepts one label from the collection of labels re-
information from one user is spread to others and 2) how this informa- ceived from neighbors following the listening rule. Using this
tion is processed by others in the dynamic process. In this method, a rule, it selects the most common label from what it observed in
Speaker-Listener based information Propagation Process (SLPP) (Xie et this step.
al., 2011) is employed to mimic human communication behavior. • The post-processing based on the labels in the memories of users is
In this respect, users are represented as a weighted graph of 𝐺 = employed to generate the communities.
(𝑉 , 𝐸, 𝑊 ), where 𝑉 is the set of nodes, 𝐸 is the set of edges denoting the
relationships between the nodes, 𝑊 is a matrix indicating the weights SLPP employs an asynchronous update scheme, so that some neigh-
of the edges, and ∣ 𝑉 ∣ is the number of nodes. Moreover, a user ratings- bors who have been updated already have memories of size 𝑡 and other
based similarity measure is used to construct the weight matrix 𝑊 . Let neighbors still have memories of size 𝑡 − 1. If the memory size is limited
𝑤𝑢,𝑣 ∈ 𝑊 be the user time-aware ratings-based similarity value between to one and the convergence criterion is met, then SLPP becomes LPA.
users 𝑢 and 𝑣, which is calculated as follows: Once the maximum number of iterations 𝑇 is reached, the SLPA stops.
The overall framework of SLPA overlapping community detection is
𝑤𝑢,𝑣 = (1) summarized in Algorithm 1.
∑ (( ) ( ) )
𝑟𝑖 (𝑢) − 𝑟̄ (𝑢) × 𝑟𝑖 (𝑣) − 𝑟̄ (𝑣) × 𝑇 𝑊 (𝑢,𝑣,𝑖)
𝑖∈𝐴𝑢,𝑣 3.4. User community-food group tendency
√
∑ (( )2 ) √ ∑ (( )2 ),
𝑟𝑖 (𝑢) − 𝑟̄ (𝑢) × 𝑇 𝑊 (𝑢,𝑣,𝑖) 𝑟𝑖 (𝑣) − 𝑟̄ (𝑣) × 𝑇 𝑊 (𝑢,𝑣,𝑖) Each individual user may have different preferences for different
𝑖∈𝐴𝑢,𝑣 𝑖∈𝐴𝑢,𝑣
food groups, regardless of which community of users he/she belongs
where 𝑟𝑖 (𝑢) ∈ 𝑅 is the rating given to food 𝑓𝑖 by user 𝑢, and 𝑟̄ (𝑢) is to. At this stage, we calculate the user community’s tendency towards
the average rating given by user 𝑢, and 𝐴𝑢,𝑣 is the set of foods which each food group (category). The tendency measure of each community
are rated by both users u and v. Moreover, 𝑇 𝑊 (𝑢,𝑣,𝑖) denotes the Time of users for each group of foods will be in the same range of recorded
Weight of the ratings of users’ rates 𝑢 and 𝑣 to food 𝑓𝑖 that calculated rates. For example, if ratings range between 1 and 5, then the tendency
as follows: measure for each community in relation to each food group will range
√ from 1 to 5 as well. More formally, the tendency measure of user com-
𝑇 𝑊 (𝑢,𝑣,𝑖) = 𝑒−𝜆(𝑇 𝑃 −𝑡(𝑢,𝑖)) × 𝑒−𝜆(𝑇 𝑃 −𝑡(𝑣,𝑖)) , (2) munity 𝑢𝑐𝑖 to food group 𝑓 𝑔𝑗 is calculated as bellow:

7
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

Algorithm 1 SLPA Overlapping Community Detection Then, for each food 𝑓𝑖 ∈ 𝐹𝑢 and each food 𝑓𝑗 ∈ 𝐹𝐶𝑜𝑚(𝑢)
′ the confidence
Input: User Graph 𝐺 = (𝑉 , 𝐸, 𝑊 ), T. value of rule 𝑓𝑖 → 𝑓𝑗 can be calculated using the following equation:
1: Step 1: Internalization
2: for i=1 to N do ( ) 𝑛(𝑓𝑖 , 𝑓𝑗 )
3: Nodes(i).Mem=i 𝑐𝑜𝑛𝑓 𝑓𝑖 → 𝑓𝑗 = , (6)
4: end for
𝑛(𝑓𝑖 )
5: Step 2: Evaluation where 𝑛(𝑓𝑖 ) is the number of users in 𝐶𝑜𝑚(𝑢) who have rated food 𝑓𝑖 ,
6: for t=1 to T do
and 𝑛(𝑓𝑖 , 𝑓𝑗 ) is the number of users in 𝐶𝑜𝑚(𝑢) who have rated both food
7: Nodes.ShuffleOrder()
8: for i=1 to N do 𝑓𝑖 and 𝑓𝑗 .
9: Listener=Nodes(i) Then, the preference rank of food 𝑓𝑗 ∈ 𝐹𝐶𝑜𝑚(𝑢)
′ for the target user 𝑢
10: Speakers=Nodes(i).getNbs() can be defined as below:
11: for j=1 to Speakers.len do
12: LabelList(j)= Speakers(j).speakerRule() ( )
13: end for
𝑃 𝑟𝑒𝑓 𝑅𝑎𝑛𝑘𝑓𝑗 (𝑢) = arg max (𝑐𝑜𝑛𝑓 𝑓𝑖 → 𝑓𝑗 × 𝑇 (𝐶𝑜𝑚(𝑢), 𝐺𝑟𝑜𝑢𝑝(𝑓𝑗 ))), (7)
𝑓𝑖 ∈𝐹𝑢
14: w=Listener.listenerRule(LabelList)
15: Listener.Mem.add(w) where 𝑇 (𝐶𝑜𝑚(𝑢), 𝐺𝑟𝑜𝑢𝑝(𝑓𝑗 )) indicates the tendency measure of user com-
16: end for munity 𝐶𝑜𝑚(𝑢) to food group 𝐺𝑟𝑜𝑢𝑝(𝑓𝑗 ) that calculated using Eq. (3).
17: end for Utilizing the Top-recommendation strategy, a recommendation list
18: Step 3: Post-processing
19: for i=1 to N do
can be generated for the target user after ranking the foods. Moreover,
20: for each food in the Top-recommendation food list and considering the
remove Nodes(i) labels seen with probability < r; ( )
21: end for rules 𝑓𝑖 → 𝑓𝑗 according to Eq (7) that leads to this recommendation
the following explanation is displayed to the target user:
∑ ( ) Food 𝑓𝑗 is recommended to you because users who have liked food
𝑢∈𝑢𝑐 𝑖 𝑇 𝑒𝑛 𝑢, 𝑓 𝑔 𝑗
𝑇 (𝑢𝑐𝑖 , 𝑓 𝑔𝑗 ) = , (3) 𝑓𝑖 and have the same food style as you are also interested in food 𝑓𝑗 .
∣ 𝑢𝑐𝑖 ∣ The preceding provides a basis for explainability in the sense that
where ∣ 𝑢𝑐𝑖 ∣ indicates the number of users in the user community 𝑢𝑐𝑖 it enables the user to understand why the given recommendation has
and 𝑇 𝑒𝑛(𝑢, 𝑓 𝑔𝑗 ) is the tendency measure of user 𝑢 to food group 𝑓 𝑔𝑗 been generated by the recommender-system.
that can be calculated as follows:
∑ 4. Experimental results
𝑟𝑓 ∈𝑅𝐹 𝑢 𝑟𝑢,𝑖
𝑇 𝑒𝑛(𝑢, 𝑓 𝑔𝑗 ) = , (4)
∣ 𝑅𝐹𝑢 𝑐 ∣
In this section, we conduct extensive experiments to demonstrate
where 𝑟𝑢,𝑖 indicates the rate given to food 𝑓𝑟𝑓 by user 𝑢 and 𝑅𝐹𝑢 is set the effectiveness of the developed Explainable Food Recommendation
of foods rated by user 𝑢. model based on a real dataset, then present the results and discuss them
in terms of various evaluation metrics. As baselines, several state-of-the-
3.5. Associative rule mining art food recommendation approaches are selected and their main ideas
are briefly described as below:
This step develops a novel user community and food group-aware
rule mining method for final explainable food recommendations. More • HAFR (Gao et al., 2019): A hierarchical attention network is used
formally, consider a set of transactions, where each transaction is a set in this model to develop a food recommendation system, which
of foods, an association rule is a rule of form 𝑋 → 𝑌 where 𝑋 and 𝑌 are considers user preferences, food ingredients, and recipe images.
sets of foods (also called food sets). This rule states that the appearance • FGCN (Gao et al., 2022): This method uses a graph convolutional
of 𝑋 in a transaction implies the presence of 𝑌 in that transaction as network to propagate information deeply within three networks
well. An example of an association rule in the user tastes analysis field based on the relations between ingredients, recipes, and users.
is: “80% of transactions that include Shrimp and Lobster also include • HGAT (Tian et al., 2022): The method uses a hierarchical graph at-
Crab; 20% of all transactions include the three of these foods”. There- tention network and exploits information about the ingredients and
fore, 𝑋 = {𝑆ℎ𝑟𝑖𝑚𝑝, 𝐿𝑜𝑏𝑠𝑡𝑒𝑟}, 𝑌 = {𝐶𝑟𝑎𝑏}, 80% stands for the confidence the user’s preferences to provide food recommendations. Further-
of this rule, and 20% corresponds to the amount of support of this rule. more, to optimize the food recommendation approach, a ranking-
A rule’s confidence indicates its degree of correlation between food sets, based objective function is introduced.
while its support indicates its significance. Originally, association rules • TDLGC (Rostami et al., 2022): This model develops a Time-Aware
were developed in order to make use of massive amounts of the rated Food Recommender-System Based on Deep Learning and Graph
items. By using this method, all significant associations between items Clustering. A food content-based recommendation system and a
will be discovered. user-based recommendation system are included in the developed
In association rule mining, the goal is to find all association rules system. In the first phase, graph clustering is used, and in the sec-
that have more than the minimum support or confidence specified by ond phase, deep-learning based user clustering is used.
the user. In other words, at this stage, the goal is to extract rules in the
form of 𝑓𝑖 → 𝑓𝑗 , so that if a user tasted (or rated) food 𝑓𝑖 previously, 4.1. Dataset
he/she will also likely be interested in food 𝑓𝑗 . Using the detected over-
lapping communities, we can extract these rules. Based on the foods During the evaluation process, we crawled a dataset from the
that have been rated by community’s members, a set of rules can be www.allrecipes.com food social network and extracted 52,821 recipes for
extracted for each community. Let 𝐹𝑢 be the set of foods rated by the the period 2000-2018. Accordingly, users’ ratings, food nutrition, and
target user 𝑢, 𝐶𝑜𝑚(𝑢) be users’ community that user 𝑢 belongs to and timestamps are crawled for each food. Foods are rated by users from
𝐺𝑟𝑜𝑢𝑝(𝑓𝑗 ) be food group that the food 𝑓𝑗 belongs to. Moreover, suppose ranges of 1 to 5 corresponding to “Couldn’t eat it”, “Didn’t like it”, “It
𝐹𝐶𝑜𝑚(𝑢) is the set of foods rated the user member of 𝐶𝑜𝑚(𝑢). Next, the was OK”, “Liked it” and “Loved it”, respectively. By rating a variety of
set of foods rated by the users in 𝐶𝑜𝑚(𝑢) but the target user 𝑢 has noted foods, implicit feedback is generated that indicates whether users in-
rated them is defined as below: teracted with them or not. We obtained in total 68,768 users, 45,630
foods, and 1,093,845 ratings after prepossessing the crawled dataset. In
𝐹𝐶𝑜𝑚(𝑢)
′
= 𝐹𝐶𝑜𝑚(𝑢) − (𝐹𝑢 ∩ 𝐹𝐶𝑜𝑚(𝑢) ). (5) addition, food images are crawled for each food item.

8
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

4.2. Evaluation measures

To evaluate the recommendation results of the compared models,

five evaluation metrics are used: Precision, Recall, F1, Normalized Dis-
counted Cumulative Gain (NDCG), and AUC. By taking into account
both the proportion of recommendations that are relevant and the pro-
portion of all relevant items in the recommendations list, precision and
recall scores are evaluated. Specifically, the following equations calcu-
late these two measures:
𝑅𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑒𝑑𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡
𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = . (8)
𝐿
𝑅𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑒𝑑𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡
𝑅𝑒𝑐𝑎𝑙𝑙 = , (9)
𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡
where 𝑅𝑒𝑐𝑜𝑚𝑚𝑒𝑛𝑑𝑒𝑑𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡 and 𝑅𝑒𝑙𝑒𝑣𝑎𝑛𝑡 are the number of recom-
mended foods that are relevant and number of all relevant foods. Fig. 4. Sensitivity analysis of 𝜆 parameter.
Moreover, the length of the recommendations list is indicated by 𝐿.
Furthermore, the F1 measure defines the harmonic mean of the preci- Table 2
sion and recalls measures as below: Performance of compared food recommendations.
2 × 𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 × 𝑅𝑒𝑐𝑎𝑙𝑙 Metrics Recommendation Models
𝐹1 = . (10)
𝑃 𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙 HAFR FGCN HGAT TDLGC EFRDIC
Moreover, another important evaluation metric focuses on the abil- Precision L=10 0.0692 0.0673 0.0703 0.0723 0.0775
ity of recommendation models to provide a list of recommendations L=15 0.0511 0.0574 0.0579 0.0649 0.0712
in which relevant items are placed at the top. This is quantified us- L=20 0.0407 0.0489 0.0531 0.0558 0.0661
ing NDCG (Normalized Discounted Cumulative Gain) measure, which Recall L=10 0.0671 0.0647 0.0682 0.0741 0.0795
L=15 0.0852 0.0882 0.0894 0.0924 0.0962
is calculated using the following equation:
L=20 0.1019 0.1061 0.1102 0.1172 0.1264
𝐷𝐶𝐺 F1 L=10 0.0681 0.066 0.0692 0.0732 0.0785
𝑁𝐷𝐶𝐺 = , (11) L=15 0.0639 0.0695 0.0703 0.0762 0.0818
𝐷𝐶𝐺𝑚𝑎𝑥
L=20 0.0582 0.0669 0.0717 0.0756 0.0868
where, NDCG L=10 0.0451 0.0437 0.0464 0.0496 0.0575
L=15 0.0492 0.0551 0.0611 0.0619 0.0689
𝐿
∑ 𝑟𝑒𝑙𝑖 L=20 0.0514 0.0591 0.0715 0.0744 0.0776
𝐷𝐶𝐺 = 𝑟𝑒𝑙1 + , (12)
𝑖=2
log2 (𝑖 + 1)

where 𝑟𝑒𝑙𝑖 is used to determine whether the recommended food 𝑓𝑖 is Comparing EFRDIC to FGCN shows the great capability of the devel-
relevant. Therefore, if the recommended food 𝑓𝑖 is relevant, then 𝑟𝑒𝑙𝑖 = oped system to produce accurate recommendations. A major factor in
1; otherwise, 𝑟𝑒𝑙𝑖 = 0. Also, 𝐷𝐶𝐺𝑚𝑎𝑥 is calculated as follows: achieving this goal is attributed to the use of user communities and
𝐿 food groups simultaneously in our EFRDIC model and also consider-
∑ 1
𝐷𝐶𝐺𝑚𝑎𝑥 = 1 + . (13) ing the temporal information of ratings in food rating prediction, which
𝑖=2
log2 (𝑖 + 1) resulted in accurate and precise recommendations.
Finally, a fourth evaluation metric is the Area Under Curve (AUC).
The AUC indicates the likelihood that a system will rank a randomly 4.4. Ablation study
chosen relevant items higher than a randomly chosen irrelevant items.
The purpose of this subsection is to design and analyze an ablation
4.3. Performance comparison study to evaluate the contribution provided by the proposed EFRDIC
model. It should be noted that EFRDIC contributes three main contribu-
In this subsection, we compare the developed EFRDIC model to tions: defining a time-aware user similarity measurement, employing an
other recommendation models. Due to the fact that the evaluation mea- overlapping community detection algorithm for user clustering, and de-
sures are computed based on the length of the recommendations list veloping a deep image clustering-based algorithm to group food items.
(L), we consider different values of this parameter (i.e., L=10, 15, 20) Through each provided contribution, the performed ablation study can
to make a better comparison. demonstrate the amount of performance improvement achieved. This
There is one input parameter in the developed EFRDIC model, 𝜆, enables us to investigate the effects of each contribution on the perfor-
that is used in Eq. (2) as a predefined threshold to determine the effect mance improvement by ignoring other factors in the proposed EFRDIC
of time factor in measuring the time-aware similarity values. The sensi- method and comparing the obtained model with the original version.
tivity analysis of the parameter 𝜆 is illustrated in Fig. 4 where we set its Fig. 5 presents the results of comparing the original version of
value from 𝜆 = 0 to 𝜆 = 4 with the step size of 0.5. As we can see from EFRDIC with the model obtained by ignoring the time-aware user sim-
these results, the best performance has been achieved when we set this ilarity measurement. According to these results, time-aware similarity
parameter to 2. measures are able to improve EFRDIC’s performance in terms of all eval-
Experimental results for compared recommendation models are uation measures as well as different lengths of recommendation lists.
summarized in Table 2 in terms of precision, recall, the F1 statistic, Hence, the time-aware similarity measurement contributes to accuracy
and the NDCG. measures through EFRDIC. This study showed that one of the main con-
Analyzing the results reported in Table 2, it is clear that the proposed tributions of the proposed time-aware similarity function (Eq. (1)) is
EFRDIC method outperforms other models on all accuracy metrics and that it contributes positively to EFRDIC efficiency. As a result, EFRDIC
differing lengths of recommendation lists. FGCN system is ranked the with time factor outperforms the model without time-factor by obtain-
second-best by achieving more promising results than other food rec- ing better results, demonstrating the positive impact of the time factor
ommendation systems that are compared to the developed system. on performance improvements. These results are expected because user

9
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

Fig. 5. The comparison of the developed system with the model was obtained by ignoring the time-aware similarity measure.

Fig. 6. The comparison of the developed system with the model was obtained by replacing the overlapping community detection.

preferences usually change over time, and taking into account this dy- neighbors and non overlapping Louvain community detection algorithm
namism results in more accurate recommendation procedures. (Blondel et al., 2008) as its prediction mechanism. From these results,
As another contribution, this study developed an overlapping com- it is clear that the original version of EFRDIC (i.e., the model with over-
munity detection algorithm, which identifies the nearest neighboring
lapping community detection) is more accurate than the other models
users to identify appropriate user communities. In order to determine
without overlapping community detection. Accordingly, the developed
whether this approach has any impact on improving the performance
method of overlapping community detection is a considerable contri-
of EFRDIC, experiments were conducted and their results are shown in
Fig. 6. These experiments compare the original version of EFRDIC (i.e., bution of this paper to improve the efficiency and effectiveness of food
with overlapping community detection) with another version that does recommendation systems. Furthermore, it is concluded that members of
not apply the community detection approach but instead uses pure K- the same community have mostly the same food style.

10
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

Fig. 7. The comparison of the developed system with the model was obtained by ignoring the deep image representation.

A comparison of the developed deep image clustering algorithm and respond the research question of RQ1, is to design a novel image-
the recommendation model obtained by ignoring the deep convolu- aware food recommendation system that takes into account food
tional feature (i.e., Original K-means) is shown in Fig. 7. Based on these images that are highly relevant to food recommendations. Taking
results, it can be demonstrated that including the deep feature encod- this information into account and detecting food groups using deep
ing in the proposed image clustering has a positive effect on boosting image clustering will allow a better recommendation process com-
the efficiency of EFRDIC model. Our developed model with deep fea- pared to previous models.
ture encoding outperforms the model without deep feature extraction, • Since none of the previously developed food recommendation sys-
in terms of performance metrics results showing that deep represen- tems could explain their recommendations, users may be discour-
tation plays a positive role in performance improvement. The results aged from following future suggestions. To respond the research
showed that adding a deep learning-based representation step to the question of RQ2 (i.e., How can the recommendation of the de-
original k-means clustering improved precision, recall, F1 and NDCG veloped recommender system be justified to user?) an explainable
by about 13.41%, 14.25%, 7.33%, and 11.32%, respectively. food recommendation method based on rule mining is developed
As a result of the above-mentioned contributions, some experiments in this paper. Accordingly, this preceding gives the user a basis for
are conducted in terms of different evaluation metrics. We can see explaining why the recommender system generated the given rec-
from these results, the highest improved performance is obtained by ommendation. This is the first work to incorporate explainability
the contribution of developed Deep image clustering in the recom- into food recommendation systems. This expects to enhance users’
mendation process. It was found that, on average, the performance trust due to the system explicit accounting of users’ communities
improved by 5.52%, 6.81%, and 11.34%, respectively, when the Time- and preferences.
aware similarity measure, Overlapping community detection, and Deep • Most food recommendation models, such as HAFR (Gao et al.,
learning-based image clustering were taken into account rather than 2019), FGCN (Gao et al., 2022) and HGAT (Tian et al., 2022),
when they were ignored. Therefore, we can conclude that the contri- have not attempted to detect user communities, and analyses of
bution of Deep learning-based image clustering plays a vital role in user-user interactions are generally light. As opposed to previous
the proposed food recommendation system and impacts the quality of works that did not take users’ communities into consideration when
recommendations significantly. These results are expectable because recommending foods and especially responding to the RQ3, the in-
ignoring the deep learning-based image clustering reduces the User tention of this work is to detect users’ communities in order to
Community-Food Group Tendency and Associative Rule Mining. define the natural groups of users. Users in the community to the
target user belongs are considered in the rule mining and the food
4.5. Discussion recommendation process (Eq. (7)). In comparison with the other
models, only the TDLGC (Rostami et al., 2022) model incorporates
Our food recommendation system utilizes key innovations that make information about users’ communities into the final recommenda-
it more effective than other state-of-the-art food recommendation sys- tion. Even though TDLGC food recommender system is superior to
tems. These innovations are as follows: other methods that have not attempted to detect the user commu-
nity, this system also ignored the advantage of overlapping user
• Despite the fact that food images convey more than just ingredients communities, which can be very common in food social networks.
and are crucial in influencing a user’s opinions about food, many Based on the ablation study of overlapping community detection
previous recommender systems, such as TDLGC (Rostami et al., method, the developed recommender system has improved in ac-
2022) and HGAT (Tian et al., 2022) ignore this information in their curacy by about 10.52% as a result of the overlapping community
food recommendation processes. The aim of this study, and also detection.

11
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

• Previous compared food recommendation models, such as HAFR recent years to predict and/or guide people’s choices based on some
(Gao et al., 2019), FGCN (Gao et al., 2022) and HGAT (Tian et predetermined criteria. While previous food recommender systems typ-
al., 2022), ignored the time factor of the historical ratings. Users’ ically achieved good results by cataloging historical interactions with
preferences, including their diets and tastes, change over time, foods and recipes to learn user’s preferences, they suffer from two
and hence a highly effective food recommendation system should major limitations. One of them is ignoring food images in the food
take this into consideration. Consequently, the previous food rec- recommendation process and lack of explainability that answer why
ommendation systems ignore the change in user preferences over specific recommendation is generated. This paper proposes a new food
time, making them ineffective. This study develops a new time- recommendation system called Explainable Food Recommendation by
aware similarity function to account for changes in preferences Deep Image Clustering, in short EFRDIC, to tackle the above limita-
over time (Eq. (1)). Thus, our developed model outperforms the tions. EFRDIC has two simultaneous steps: (1) Deep learning-based
state-of-the-art models, which often ignore the time factor. Accord- food image clustering, and (2) User community detection, followed by
ing to the results of the ablation study shown in Fig. 5, considering two other steps: (3) User community-food group tendency estimation,
the time factor significantly enhanced the performance of the pro- (4) Explainability module using associative rule mining. The proposed
posed EFRDIC model. method is evaluated through several experiments in comparison with
state-of-the-art food recommendation methods. In terms of precision,
As can be seen from the results section, our developed model al- recall, F1, and NDCG metrics, the proposed method clearly outperforms
ready outperforms several state-of-the-art food recommendation mod- other models. Furthermore, an ablation study is conducted to test the
els. However, the question remains whether it can be further improved. effectiveness of each contribution to the proposed method. According
The process of our model is hampered by several factors that prevent to the ablation study, the contributions significantly enhanced the ef-
further improvement: fectiveness of the proposed food recommendation model.
The paper opens up new door for testing the developed model in
• Our developed system cannot guarantee that the recommended other culture-specific food dataset, which still to be obtained or elicited
foods are healthy choice for the users, since this study do not incor- from appropriate sources.
porate nutrition information of the foods. In fact, the vast majority
of previous works focused only on the user’s preferences. There- CRediT authorship contribution statement
fore, for future works we plan to incorporate, health and nutrition
factors into the food recommendation framework, so that the rec- Mehrdad Rostami: Conceptualization, Methodology, Software,
ommender system guides users to healthy eating style. In other Writing – Reviewing and Editing. Usman Muhammad: Methodology,
words, we aim to focus on presenting the general architecture for Software. Saman Forouzandeh: Visualization, Investigation. Kamal
implementing the food recommendation system based on prefer- Berahmand: Visualization, Investigation. Vahid Farrahi: Supervision,
ence and nutritional information. Writing – Reviewing and Editing. Mourad Oussalah: Supervision, Writ-
• Because publicly accessible food datasets, such as Food-101 ing – Reviewing and Editing.
(Bossard et al., 2014) and Yummly (Min et al., 2016), do not
provide relevant user-rated information, they cannot be used to Declaration of competing interest
evaluate food recommendation systems. To obtain the user-food
rating dataset, we crawled Allrecipes food social network. In spite The authors declare that they have no known competing financial
of the fact that this website is one of the largest food social net-
interests or personal relationships that could have appeared to influence
works, according to our analysis, 85% of its visitors of this food
the work reported in this paper.
social network are North Americans who have a special food cul-
ture. Therefore, evaluating our recommended system using the data
Data availability
from sites hosted by countries with different food cultures could
introduce some bias into the results. Therefore, we plan to crawl
Data will be made available on request.
other social food networks as well as analyze diet styles in different
cultures and countries to overcome this shortcoming.
Acknowledgements
• Despite the fact that the explainable food recommender system at-
tempts to convince users to try other favorite foods, it does not
provide clear measures of how effective the explanations are in The project is supported by the University of Oulu Academy of
influencing users. Therefore, novel performance metrics could be Finland Profi5 on Digihealth (project number 326291). Moreover, this
introduced to evaluate the generated explanations. work also was supported in part by the Ministry of Education and Cul-
• Our developed recommender system as well as most of previous ture, Finland (OKM/20/626/2022).
ones typically ignore user characteristics (e.g. age, height, weight,
gender, location, allergies, medical history, etc.) and only consider References
food content or user ratings when recommending foods to a user.
Aditya, G. M., Hoode, A., Rai, K. A., Biradar, G., Kumara, M. A. A., Kumar, M. V. M.,
Since each person’s nutritional requirements can substantially dif-
Prashanth, B. S., Sneha, H. R., & Shivadarshan, S. L. (2021). Machine learning based
fer according to user’s profile, an efficient food recommender sys- platform and recommendation system for food ordering services within premises. In
tem should take their user characteristics into account when gener- 2021 2nd global conference for advancement in technology (GCAT) (pp. 1–8). IEEE.
ating recommendations. Future research will involve incorporating Agapito, G., Simeoni, M., Calabrese, B., Guzzi, P. H., Fuiano, G., & Cannataro, M. (2017).
additional side information about users into the food recommenda- Dietos: A recommender system for health profiling and diet management in chronic
diseases. In HealthRecSys@ RecSys (pp. 32–35).
tion framework.
Ahmadian, S., Joorabloo, N., Jalili, M., & Ahmadian, M. (2022). Alleviating data sparsity
problem in time-aware recommender systems using a reliable rating profile enrich-
5. Conclusion ment approach. Expert Systems with Applications, 187, Article 115849.
Albatayneh, N. A., Ghauth, K. I., & Chua, F.-F. (2022). Discriminate2rec: Negation-
Designing intelligent tools to find the tastes of users regarding avail- based dynamic discriminative interest-based preference learning for semantics-aware
content-based recommendation. Expert Systems with Applications, 199, Article 116988.
able foods and assist them in consuming a more balanced and healthy Ali, F., Islam, S. M. R., Kwak, D., Khan, P., Ullah, N., Yoo, S.-j., & Kwak, K. S. (2018). Type-
diet is one of the challenges in the field of recommendation systems. 2 fuzzy ontology–aided recommendation systems for iot–based healthcare. Computer
There have been several food recommendation systems developed in Communications, 119, 138–155.

12
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

Bishop, T. R. P., von Hinke, S., Hollingsworth, B., Lake, A. A., Brown, H., & Burgoine, T. Kunaver, M., & Požrl, T. (2017). Diversity in recommender systems – a survey. Knowledge-
(2021). Automatic classification of takeaway food outlet cuisine type using machine Based Systems, 123, 154–162.
(deep) learning. Machine Learning with Applications, 6, Article 100106. Lee, H. I., Choi, I. Y., Moon, H. S., & Kim, J. K. (2020). A multi-period product recom-
Blondel, V. D., Guillaume, J.-L., Lambiotte, R., & Lefebvre, E. (2008). Fast unfolding of mender system in online food market based on recurrent neural networks. Sustain-
communities in large networks. Journal of Statistical Mechanics: Theory and Experi- ability, 12(3), 969.
ment, 2008(10), Article P10008. Lipton, Z. C. (2018). The mythos of model interpretability: In machine learning, the con-
Bossard, L., Guillaumin, M., & Van Gool, L. (2014). Food-101–mining discriminative com- cept of interpretability is both important and slippery. Queue, 16(3), 31–57.
ponents with random forests. In European conference on computer vision (pp. 446–461). Liu, N., Ge, Y., Li, L., Hu, X., Chen, R., & Choi, S.-H. (2020). Explainable recommender
Springer. systems via resolving learning representations. In Proceedings of the 29th ACM inter-
Chavan, P., Thoms, B., & Isaacs, J. (2021). A recommender system for healthy food national conference on information & knowledge management (pp. 895–904).
choices: Building a hybrid model for recipe recommendations using big data sets. Maia, R., & Ferreira, J. C. (2018). Context-aware food recommendation system
In Proceedings of the 54th Hawaii international conference on system sciences (p. 3774). (pp. 349–356).
Chen, C.-H., & Toumazou, C. (2019). Personalized expert recommendation systems for Min, W., Jiang, S., & Jain, R. (2019). Food recommendation: Framework, existing solu-
optimized nutrition. Trends in Personalized Nutrition, 309–338. tions, and challenges. IEEE Transactions on Multimedia, 22(10), 2659–2671.
Cheng, G., Li, Z., Yao, X., Guo, L., & Wei, Z. (2017). Remote sensing image scene classifi-
Min, W., Jiang, S., Sang, J., Wang, H., Liu, X., & Herranz, L. (2016). Being a supercook:
cation using bag of convolutional features. IEEE Geoscience and Remote Sensing Letters,
Joint food attributes and multimodal content modeling for recipe retrieval and explo-
14(10), 1735–1739.
ration. IEEE Transactions on Multimedia, 19(5), 1100–1113.
Cui, Z., Xu, X., Fei, X. U. E., Cai, X., Cao, Y., Zhang, W., & Chen, J. (2020). Personal-
Molina-Ayala, M. A., Rodríguez-Amador, V., Suárez-Sánchez, R., León-Solís, L., Gómez-
ized recommendation system based on collaborative filtering for iot scenarios. IEEE
Zamudio, J., Mendoza-Zubieta, V., Cruz, M., & Suárez-Sánchez, F. (2022). Expression
Transactions on Services Computing, 13(4), 685–695.
of obesity-and type-2 diabetes-associated genes in omental adipose tissue of individ-
Deebak, B. D., & Al-Turjman, F. (2020). A novel community-based trust aware recom-
uals with obesity. Gene, Article 146181.
mender systems for big data cloud service networks. Sustainable Cities and Society, 61,
Naik, P. A. (2020). Intelligent food recommendation system using machine learning. Nu-
Article 102274.
trition, 5(8).
Ding, S., Feng, F., He, X., Liao, Y., Shi, J., & Zhang, Y. (2022). Causal incremental graph
convolution for recommender system retraining. IEEE Transactions on Neural Networks Norouzi, S., Nematy, M., Zabolinezhad, H., Sistani, S., & Etminani, K. (2017). Food recom-
and Learning Systems. mender systems for diabetic patients: A narrative review. Reviews in Clinical Medicine,
Forouzandeh, S., Berahmand, K., Nasiri, E., & Rostami, M. (2021). A hotel recommender 4(3), 128–130.
system for tourists using the artificial bee colony algorithm and fuzzy topsis model: Noulapeu Ngaffo, A., El Ayeb, W., & Choukair, Z. (2021). A time-aware service recommen-
A case study of tripadvisor. International Journal of Information Technology & Decision dation based on implicit trust relationships and enhanced user similarities. Journal of
Making, 20(01), 399–429. Ambient Intelligence and Humanized Computing, 12(2), 3017–3035.
Forouzandeh, S., Berahmand, K., & Rostami, M. (2021). Presentation of a recommender Oh, Y., Choi, A., & Woo, W. (2010). u-babsang: A context-aware food recommendation
system with ensemble learning and graph embedding: A case on movielens. Multime- system. Journal of Supercomputing, 54(1), 61–81.
dia Tools and Applications, 80(5), 7805–7832. Peake, G., & Wang, J. (2018). Explanation mining: Post hoc interpretability of latent
Forouzandeh, S., Rostami, M., & Berahmand, K. (2021). Presentation a trust walker for factor models for recommendation systems. In Proceedings of the 24th ACM SIGKDD
rating prediction in recommender system with biased random walk: Effects of h- international conference on knowledge discovery & data mining (pp. 2060–2069).
index centrality, similarity in items and friends. Engineering Applications of Artificial Premasundari, M., & Yamini, C. (2019). Food and therapy recommendation system for
Intelligence, 104, Article 104325. autistic syndrome using machine learning techniques. In 2019 IEEE international con-
Forouzandeh, S., Rostami, M., & Berahmand, K. (2022). A hybrid method for recommen- ference on electrical, computer and communication technologies (ICECCT) (pp. 1–6).
dation systems based on tourism with an evolutionary algorithm and topsis model. IEEE.
Fuzzy Information and Engineering, 1–25. Princy, J., Senith, S., Kirubaraj, A. A., & Vijaykumar, P. (2021). A personalized food
Gao, X., Feng, F., He, X., Huang, H., Guan, X., Feng, C., Ming, Z., & Chua, T.-S. (2019). recommender system for women considering nutritional information. International
Hierarchical attention network for visually-aware food recommendation. IEEE Trans- Journal of Pharmaceutical Research, 13(2).
actions on Multimedia, 22(6), 1647–1659. Rehman, F., Khalid, O., Bilal, K., Madani, S. A., et al. (2017). Diet-right: A smart food
Gao, X., Feng, F., Huang, H., Mao, X.-L., Lan, T., & Chi, Z. (2022). Food recommendation recommendation system. KSII Transactions on Internet and Information Systems (TIIS),
with graph convolutional network. Information Sciences, 584, 170–183. 11(6), 2910–2925.
Ge, M., Ricci, F., & Massimo, D. (2015). Health-aware food recommender system. In Pro- Rostami, M., Oussalah, M., & Farrahi, V. (2022). A novel time-aware food recommender-
ceedings of the 9th ACM conference on recommender systems (pp. 333–334). system based on deep learning and graph clustering. IEEE Access.
Gusnedi, G., Fahmida, U., Witjaksono, F., Nurwidya, F., Mansyur, M., Djuwita, R., Dwiri- Samad, S., Ahmed, F., Naher, S., Kabir, M. A., Das, A., Amin, S., & Islam, S. M. S. (2022).
ani, C. M., & Abdullah, M. (2022). Effectiveness of optimized food-based recommen- Smartphone apps for tracking food consumption and recommendations: Evaluating
dation promotion to improve nutritional status and lipid profiles among minangkabau artificial intelligence-based functionalities, features and quality of current apps. Intel-
women with dyslipidemia: A cluster-randomized trial. BMC Public Health, 22(1), ligent Systems with Applications, Article 200103.
1–12. Sánchez-Moreno, D., Zheng, Y., & Moreno-García, M. N. (2020). Time-aware music rec-
Halim, Z., Ali, O., & Khan, M. G. (2019). On the efficient representation of datasets as ommender systems: Modeling the evolution of implicit user preferences and user
graphs to mine maximal frequent itemsets. IEEE Transactions on Knowledge and Data listening habits in a collaborative filtering approach. Applied Sciences, 10(15), 5324.
Engineering, 33(4), 1674–1691.
Shimizu, R., Matsutani, M., & Goto, M. (2022). An explainable recommendation frame-
Halim, Z., & Rehan, M. (2020). On identification of driving-induced stress using elec-
work based on an improved knowledge graph attention network with massive vol-
troencephalogram signals: A framework based on wearable safety-critical scheme and
umes of side information. Knowledge-Based Systems, 239, Article 107970.
machine learning. Information Fusion, 53, 66–79.
Sookrah, R., Dhowtal, J. D., & Nagowah, S. D. (2019). A dash diet recommendation system
Halim, Z., Sulaiman, M., Waqas, M., & Aydın, D. (2022). Deep neural network-based iden-
for hypertensive patients using machine learning. In 2019 7th international conference
tification of driving risk utilizing driver dependent vehicle driving features: A scheme
on information and communication technology (ICoICT) (pp. 1–6). IEEE.
for critical infrastructure protection. Journal of Ambient Intelligence and Humanized
Srilakshmi, M., Chowdhury, G., & Sarkar, S. (2022). Two-stage system using item fea-
Computing, 1–19.
tures for next-item recommendation. Intelligent Systems with Applications, 14, Arti-
Hiriyannaiah, S., Siddesh, G. M., & Srinivasa, K. G. (2020). Deep visual ensemble sim-
cle 200070.
ilarity (dvesm) approach for visually aware recommendation and search in smart
Starke, A. D., & Trattner, C. (2021). Promoting healthy food choices online: A case for
community. Journal of King Saud University: Computer and Information Sciences.
multi-list recommender systems. In Proceedings of the ACM IUI 2021 workshops.
James, S. L., Abate, D., Abate, K. H., Abay, S. M., Abbafati, C., Abbasi, N., Abbastabar, H.,
Abd-Allah, F., Abdela, J., Abdelalim, A., et al. (2018). Global, regional, and national Subramaniyaswamy, V., Manogaran, G., Logesh, R., Vijayakumar, V., Chilamkurti, N.,
incidence, prevalence, and years lived with disability for 354 diseases and injuries for 195 Malathi, D., & Senthilselvan, N. (2019). An ontology-driven personalized food rec-
countries and territories, 1990–2017: A systematic analysis for the global burden of disease ommendation in iot-based healthcare system. Journal of Supercomputing, 75(6),
study 2017. The Lancet, 392(10159), 1789–1858. 3184–3216.
Kang, S., Lee, D., Kweon, W., & Yu, H. (2022). Personalized knowledge distillation for Sulthana, A. R., Gupta, M., Subramanian, S., & Mirza, S. (2020). Improvising the perfor-
recommender system. Knowledge-Based Systems, 239, Article 107958. mance of image-based recommendation system using convolution neural networks
Karimi, M., Jannach, D., & Jugovac, M. (2018). News recommender systems – survey and and deep learning. Soft Computing, 24(19), 14531–14544.
roads ahead. Information Processing & Management, 54(6), 1203–1227. Teng, C.-Y., Lin, Y.-R., & Adamic, L. A. (2012). Recipe recommendation using ingredient
Kefalas, P., & Manolopoulos, Y. (2017). A time-aware spatio-textual recommender system. networks. In Proceedings of the 4th annual ACM web science conference (pp. 298–307).
Expert Systems with Applications, 78, 396–406. Thongsri, N., Warintarawej, P., Chotkaew, S., & Saetang, W. (2022). Implementation of a
Kim, J.-C., & Chung, K. (2020). Knowledge-based hybrid decision model using neural personalized food recommendation system based on collaborative filtering and knap-
network for nutrition management. Information Technology and Management, 21(1), sack method. International Journal of Electrical and Computer Engineering [ISSN 2088-
29–39. 8708], 12(1).
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep Tian, Y., Zhang, C., Metoyer, R., & Chawla, N. V. (2022). Recipe recommendation with
convolutional neural networks. Advances in Neural Information Processing Systems, 25. hierarchical graph attention network. Frontiers in Big Data, 4, Article 778417.

13
M. Rostami, U. Muhammad, S. Forouzandeh et al. Intelligent Systems with Applications 16 (2022) 200157

Tran, T. N. T., Felfernig, A., Trattner, C., & Holzinger, A. (2021). Recommender systems Yu, W., He, X., Pei, J., Chen, X., Xiong, L., Liu, J., & Qin, Z. (2021). Visually aware
in the healthcare domain: State-of-the-art and research issues. Journal of Intelligent recommendation with aesthetic features. The VLDB Journal, 30(4), 495–513.
Information Systems, 57(1), 171–201. Zhang, F., Qu, Y., Xu, Y., & Wang, S. (2020). Graph embedding-based approach for de-
Trattner, C., & Elsweiler, D. (2017). Investigating the healthiness of internet-sourced tecting group shilling attacks in collaborative recommender systems. Knowledge-Based
recipes: Implications for meal planning and recommender systems. In Proceedings of Systems, 199, Article 105984.
the 26th international conference on world wide web (pp. 489–498). Zhang, H.-R., Qian, J., Qu, H.-L., & Min, F. (2022). A mixture-of-Gaussians model for
Ullah, F., Zhang, B., & Khan, R. U. (2019). Image-based service recommendation system: estimating the magic barrier of the recommender system. Applied Soft Computing,
A jpeg-coefficient rfs approach. IEEE Access, 8, 3308–3318. 114, Article 108162.
Vairale, V. S., & Shukla, S. (2019). Recommendation framework for diet and exercise Zhang, X., Luo, H., Chen, B., & Guo, G. (2020). Multi-view visual bayesian personalized
based on clinical data: A systematic review. Data Science and Big Data Analytics, ranking for restaurant recommendation. Applied Intelligence, 50(9), 2901–2915.
333–346. Zhang, Y., Chen, X., et al. (2020). Explainable recommendation: A survey and new per-
Van Dat, N., Van Toan, P., & Thanh, T. M. (2022). Solving distribution problems in spectives. Foundations and Trends in Information Retrieval, 14(1), 1–101.
content-based recommendation system with Gaussian mixture model. Applied Intel- Zhang, Y., Lai, G., Zhang, M., Zhang, Y., Liu, Y., & Ma, S. (2014). Explicit factor models for
ligence, 52(2), 1602–1614. explainable recommendation based on phrase-level sentiment analysis. In Proceedings
Viktoratos, I., Tsadiras, A., & Bassiliades, N. (2018). Combining community-based knowl- of the 37th international ACM SIGIR conference on research & development in information
edge with association rule mining to alleviate the cold start problem in context-aware retrieval (pp. 83–92).
recommender systems. Expert Systems with Applications, 101, 78–90. Zhao, J., Yang, S., Huo, H., Sun, Q., & Geng, X. (2021). Tbtf: An effective time-varying bias
Wang, D. D., Li, Y., Afshin, A., Springmann, M., Mozaffarian, D., Stampfer, M. J., Hu, F. B., tensor factorization algorithm for recommender system. Applied Intelligence, 51(7),
Murray, C. J. L., & Willett, W. C. (2019). Global improvement in dietary quality could 4933–4944.
lead to substantial reduction in premature death. The Journal of Nutrition, 149(6), Zhu, D., Toker, M., Shyr, W., Fram, E., Watts, K. L., & Agalliu, I. (2022). Association of
1065–1074. obesity and diabetes with prostate cancer risk groups in a multiethnic population.
Wang, K., Zhang, T., Xue, T., Lu, Y., & Na, S.-G. (2020). E-commerce personalized rec- Clinical Genitourinary Cancer.
ommendation analysis by deeply-learned clustering. Journal of Visual Communication
and Image Representation, 71, Article 102735.
Xie, J., Szymanski, B. K., & Liu, X. (2011). Slpa: Uncovering overlapping communities in
social networks via a speaker-listener interaction dynamic process. In 2011 ieee 11th
international conference on data mining workshops (pp. 344–349). IEEE.

14