Neurocomputing 576 (2024) 127326

Contents lists available at ScienceDirect

Neurocomputing
journal homepage: www.elsevier.com/locate/neucom

A novel healthy food recommendation to user groups based on a deep social

community detection approach
Mehrdad Rostami a, *, Kamal Berahmand b, Saman Forouzandeh c, Sajad Ahmadian d,
Vahid Farrahi a, e, f, Mourad Oussalah a, f
a
Center for Machine Vision and Signal Analysis (CMVS), Faculty of Information Technology and Electrical Engineering, University of Oulu, Oulu, Finland
b
School of Computer Science, Faculty of Science, Queensland University of Technology, Brisbane, Australia
c
School of Mathematics and Statistics, University of New South Wales, Sydney, NSW, Australia
d
Faculty of Information Technology, Kermanshah University of Technology, Kermanshah, Iran
e
Institute for Sport and Sport Science, TU Dortmund University, Dortmund, Germany
f
Research Unit of Health Sciences and Technology, Faculty of Medicine, University of Oulu, Oulu, Finland

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

Communicated by Zidong Wang Existing food recommendation models have typically suggested foods or recipes to single users. However, in
reality, users may be members of a group, family, or community, requiring food recommendation systems to
Keywords: support the whole group. Food recommendations to groups are a more challenging task than food recommen­
Recommender systems dations to individuals, as each person’s preferences in the group should be addressed before giving the recom­
Food recommendation
mendations. Suggesting healthy food is also important in a food recommendation system, given that unhealthy
Group recommendation
diets can lead to different diseases. To address these challenges, a new healthy group food recommendation
Community detection
Deep learning system based on deep social community detection and user popularity is developed in this study. To this end, an
innovative deep community detection approach based on feature learning and deep neural networks is developed
using the calculated time-aware user similarity measure. In addition, a health-aware rate prediction measure­
ment, which considers both group preferences and health factors, is developed. Different experiments are
designed on two real-food social networks to specify the efficiency of the suggested model, and the results
indicate that it enhanced the single-user and group satisfaction metrics.

1. Introduction recommended food aligns with their individual preferences. Therefore,

for health-oriented food services and recommendations, the food
The emergence of the Internet and web-based technologies has recommendation system should suggest food options that are in line
enabled people to access information effortlessly and instantly. There­ with the user’s preferences [8,9]. Otherwise, users will not be motivated
fore, online systems such as social media, online shopping, and lifestyle to follow the suggestions.
applications have profoundly impacted people’s daily routines [1–4]. Recent advancements in online food applications have led to the
Social media platforms allow individuals to share pictures of their meals development of numerous food recommendation systems that are
with friends, whereas lifestyle apps offer suggestions on what to eat tailored to individual user preferences [10,11]. However, several chal­
[5–7]. Among the diverse applications of lifestyle services, food lenges persist in this domain, which require special attention. Tradi­
recommendation systems have become a critical component in pro­ tional food recommendation models have focused on single-user
moting healthy living. Food recommendation systems are crucial for scenarios. However, most real-life interactions occur in groups. For
predicting user preferences, including recipes, eating habits, and instance, one often goes to a restaurant with friends, has dinner with
behavior. Therefore, for a food recommendation system to be effective, colleagues, or plans to cook a meal with their family. To facilitate these
it is essential to accurately understand the food preferences of users. This group activities, food recommendation models should recommend foods
is especially important when providing health-oriented food services, as to groups according to the interests and priorities of members instead of
users are more likely to embrace and follow a suggestion if the targeting a single individual. Understanding the challenges of group

* Corresponding author.
E-mail address: [email protected] (M. Rostami).

https://doi.org/10.1016/j.neucom.2024.127326
Received 27 March 2023; Received in revised form 15 September 2023; Accepted 21 January 2024
Available online 25 January 2024
0925-2312/© 2024 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M. Rostami et al. Neurocomputing 576 (2024) 127326

recommendation is vital in an age where collective decision-making • This research proposes an innovative health-aware rate prediction
forms the crux of many real and digital experiences. Group recom­ criterion that considers both group favorites and the health factors of
mendation introduces multi-dimensional complexity, where the task is foods. To the best of our knowledge, this is the first study that uses a
not only to cater to the diverse preferences of individual members but health-aware rate prediction mechanism in group recommendation
also to harmonize them into a coherent recommendation that maximizes algorithms. By considering both taste preferences and health aspects,
overall group satisfaction. Unlike single-user recommendations, which our proposed method can provide effective and personalized rec­
operate on a linear understanding of user behavior and preferences, ommendations for groups of users.
group recommendations must juggle conflicting interests, rank prior­
ities, and navigate the potential pitfalls of group dynamics [12]. Tradi­ The remainder of this paper is structured as follows: Section 2 dis­
tional single-user recommendation strategies, which are rooted in cusses the related works. In Section 3, we present a detailed description
singular preference mapping, fall short when faced with this layered of the suggested food recommendation model for user groups. The
challenge. The nuances of group interplay and the imperative of designed experiments for evaluation and method analysis are presented
reaching a consensus require an approach that is tailored to capturing in Section 4. Lastly, Section 5 concludes the study and provides some
the subtleties and distinctiveness of group-based interactions and de­ suggestions for future work.
cisions. Traditional individual recommendation systems are therefore
limited in their ability to suggest items to the users’ group. Therefore, 2. Related works
researchers have created different recommendation models to address
this issue. Group recommender systems can optimize the overall satis­ In this section, we provide a comprehensive review of related liter­
faction of a group by balancing the preferences of its members. In ature, categorizing the works into four main groups: food recommender
contrast, classic single recommendation models focus on meeting the systems, community detection-based models, time-aware algorithms,
needs of single users, which may not be feasible when their interests are and group recommender systems. Upon analysis, we demonstrate that
inconsistent. our recommender system is the only one that includes all four categories
On the other hand, the importance of a well-balanced diet for a in its module, making it superior to the recommenders in the existing
person’s health and well-being cannot be overstated. According to the literature.
global burden of disease research, diet style contributes remarkably to
malnutrition, overweight, and unhealthy eating habits, which cause 11 2.1. Food recommendation systems
million preventable early deaths each year [13]. Therefore, it is crucial
to understand how to offer personalized, healthy food recommenda­ Food recommendation models have attracted substantial attention in
tions. Despite the importance of healthy food recommendation systems, the last few years, especially in the context of promoting a healthy
many previous models have failed to ensure that recommended foods lifestyle [21]. The reviewed literature in the food domain has concen­
are healthy choices due to a lack of nutrition information, and they have trated on user preferences and health concerns for food recommenda­
focused on user preferences while disregarding the health factor in tions. For instance, in [16], the authors developed a novel food
recommended foods [14–16]. recommendation system that considers both food ingredients and im­
By measuring food health factors and incorporating the social and ages. The proposed method utilizes the user’s historical ratings, food
behavioral characteristics of users, we developed a novel food recom­ ingredients, and food images to provide recommendations. Moreover, a
mendation system to address the above-mentioned issues. Compared to model is introduced in [11] that suggests recipes based on ingredients
previously developed food recommendation systems, which often lack available in the refrigerator. In [22], a novel model was introduced to
time-awareness [15,17,18], rely on graph-based user clustering [14,19], extract user food preferences by considering online restaurant reviews
and overlook health considerations [14,17,20], the proposed model using NLP models. Sentiment analysis was employed to determine
offers substantial advancements in providing healthy food recommen­ whether the user’s opinions on each food item were reasonable or un­
dations for user groups. The following are our main contributions: reasonably negative. Furthermore, the authors of [23] identified that
ingredient–ingredient, ingredient–recipe, and recipe–user interactions
• This study introduces a time-aware similarity criterion to enhance are essential in food recommendation models. Therefore, they intro­
the accuracy of the user–user similarity matrix by accounting for the duced a novel food recommendation model that employs a graph con­
temporal knowledge of ratings. To achieve this, we introduce a volutional network alongside embedding propagation layers, enabling
weighting procedure that considers the significance of the time co­ enhanced representation learning and the modeling of high-order re­
efficient for ratings. Specifically, older ratings are given less weight lations. Finally, a hierarchical attention-based model was proposed in
than newer ones. By incorporating this time-aware feature into the this study to prioritize users’ preferences.
recommendation process, we can dynamically capture users’ In addition to learning users’ preferences, food recommendation
evolving preferences over time. systems should also analyze health problems, past diet styles, and
• A novel deep community detection technique using feature learning nutritional needs to promote a healthy diet [24]. For instance, in [6], a
and deep neural networks is employed to specify the natural novel Android platform-based food recommendation system that allows
groupings of users. As a result of this community detection algo­ users to choose food based on their favorites and health requirements
rithm, a trust network is created to improve the efficiency of leaders was developed. The authors of [25] assessed the health factor of recipes
and influencers in the user groups. in food social networks using nutritional measures and analyzed user
• To identify the core users in each group, a novel user popularity activities to determine their preference for healthy foods. Diet-Right, an
measure is introduced. Our developed measure incorporates node online food recommendation system based on ant colony optimization,
centrality concepts with trust networks to score user’s popularity in was developed by [26], which emphasizes nutrition when suggesting
each group. diets for patients. By employing matrix factorization and feature engi­
• A novel measure for aggregating users’ preferences is developed. As neering, a context-aware food recommendation model that lets users
part of our measure, we consider the popularity of users and their order healthy food in real-time and access nearby establishments was
food preferences to aggregate this information and form the group developed in [27]. In addition, Wang et al. [12] proposed Market2Dish,
profile. As far as our knowledge goes, this is the first group recom­ which is a healthy food recommendation system for assisting individuals
mendation to be developed in the food recommendation system in discovering personalized food options and adopting a healthy life­
research field. style, thus helping to prevent illnesses that may result from an unhealthy
diet. Lastly, to maintain healthy diets and prevent health complications,

2
M. Rostami et al. Neurocomputing 576 (2024) 127326

the authors of [28] developed a diet recommendation model. to capture these changes. Moreover, the authors of [48] developed a
All the previously mentioned food recommendation models recom­ time-aware recommendation system that infers implicit trust relation­
mend foods to users individually; however, many activities, such as ships to improve suggestion reliability by calculating both implicit and
dining out, are typically performed by people in groups. Reviewing the explicit trust connections. Furthermore, in [49], a time-aware recom­
aforementioned food recommendation models revealed that these mendation model was introduced by combining reliability and confi­
models fail to consider the group recommendation aspect because only dence measures and appraising predictability using a temporal
individual needs are considered. In this paper, this problem is efficiently reliability criterion that considers changes in user favorites over time.
addressed by proposing an innovative food group recommendation and Lastly, in [50], a stock recommendation approach was developed by
incorporating user group satisfaction into the recommendation capturing dynamic attributes from time-series data using a stacked
procedure. graph neural network model. This approach considers the dynamic
behavior of input data by recognizing the changes in patterns during
2.2. Community detection-based models different time periods.

Community detection has become a prevalent technique in recom­ 2.4. Group recommendation systems
mendation systems for understanding how network structures affect
behavior patterns [29,30]. Underlying structures can be identified In [51–53], group recommender systems for recommending items
through the detection of communities, which has been deemed a highly that are interesting, suitable, and useful to groups of people rather than
effective approach [31,32]. These models can be employed for uncov­ individuals were suggested. To identify commonly preferred items, as
ering unknown patterns, such as discovering potential friends with agreed upon by group members, innovative aggregation methods that
similar preferences on social networks. Algorithms used to detect com­ hybridize increasing function and approval voting models were pro­
munities are mainly designed to capture the tendency for nodes within a posed in [54]. The authors demonstrated the effectiveness of hybridized
graph to group based on their similarities [33,34]. Community detection methods by comparing them with previous models using three
techniques in recommendation models can improve rating prediction real-world benchmark datasets, including MovieLens Public, MovieLens
accuracy and adequately handle cold starts and sparse data. To tackle Million, and Netflix datasets. Seo et al. [55] introduced a genre prefer­
the cold start problem and improve final recommendation performance, ence definition and grouped the groups based on their definitions. Due
Paleti et al. [35] introduced a hybrid recommender system approach by to the lower number of genres versus items, this group recommendation
combining the strengths of community detection and the alternating can reduce the complexity of time more effectively. As revealed by the
least squares algorithm. The study presented a potential approach to the designed experiments, the genre-based group recommendation model
cold start issue. Specifically, the community detection technique is considerably enhances clustering time performance. Moreover, using
employed to recognize communities of users with similar preferences, the Expectation Maximization (EM) technique, Ismailoglu [56] aggre­
whereas the alternating least squares algorithm is utilized to calculate gated the favorites of group users to predict group favorites. Using this
the latent features of the users and items to enhance the final recom­ recommender system, group preferences can be estimated based on any
mendation quality. Moreover, integrating community detection and background knowledge about group members. In [57], an innovative
association rule mining was proposed in [36] to introduce a recom­ group recommendation method based on the aggregation of users’ deep
mendation model for context-aware users and item cold start issues. A profiles was developed. Generally, deep profiles refer to the profiles of
personalized recommendation model was introduced in [37] using a users that lie deep within recommendation models. The results of the
user group identification technique and a time-aware matrix to accu­ experiment indicated that this group recommender system performed
rately predict the preferences of the users. A community-aware recom­ better than the models that relied on item preferences. The authors of
mendation model was developed in [38] for cloud-based services. An [58] aimed to develop a hierarchical visualization model for group
e-commerce recommendation model utilizing deep learning-based user recommendation that would provide a visual presentation and an
grouping techniques and collaborative filtering-based approaches to intuitive explanation of group recommendation. To organize the en­
process large-scale user data was introduced in [39]. In [40], a graph tities, a hierarchy graph was used in the first step. Each entity was then
embedding-based algorithm was used to detect group shilling attacks in associated with a circular graph, with each piece representing an indi­
traditional recommendation models. Moreover, a group recommenda­ vidual user, allowing the strength of the members to be tracked. Yin
tion system using global and local structures of social networks was et al. [59] also presented an end-to-end group recommendation model
designed in [41], where users in the same community contribute based on Graph Convolutional Networks and a bipartite graph embed­
differently according to their social influence obtained by an attentive ding model as basic building blocks to learn the representations of
graph convolutional network. groups and individuals.
Notably, the aforementioned group recommendation models did not
2.3. Time-aware models consider health factors when predicting the rating of items. Their focus
was on improving the accuracy of recommendations by considering
Time-aware recommendation models have focused on modeling various performance-related factors. However, in the context of group
users’ preferences while considering temporal dynamics [42,43]. Such recommendations, especially food recommendations, accounting for
models consider that users’ favorites are prone to change over time as health factors is crucial for providing reliable and trustworthy recom­
their behavior evolves [44]. To improve the efficiency of recommender mendations. Accordingly, this paper introduces an innovative health-
systems, the authors of [43] suggested a time-aware system that utilizes aware rating prediction approach that combines both health and effi­
two unified techniques to consider users’ past ratings and item charac­ ciency factors to be used for healthy group food recommendation.
teristics and evaluates how time affects various time intervals. Similarly,
a customized recommendation model for IoT services was proposed in 3. Proposed model
[45], utilizing a collaborative filtering-based approach that integrated
the time correlation coefficient and clustering techniques. To capture This section details the proposed model, referred to as the Healthy
the evolution of implicit user preferences over time, a time-aware music Group Food Recommendation System (HGFR), based on user popularity
recommendation system that employs collaborative filtering to charac­ and food health factors and divided into five primary stages: (1) user
terize daily listening habits was developed in [46]. In [47], user bias was similarity calculation and graph representation of users, (2) deep com­
suggested as a measure to quantify changes in user favorites over time, munity detection using a generated user similarity matrix, (3) trust
and a time-varying bias tensor factorization-based method was proposed network generation, (4) user popularity calculation, (5) health-aware

3
M. Rostami et al. Neurocomputing 576 (2024) 127326

Fig. 1. Overview of the developed health group food recommendation model.

single rate prediction, and (6) user rating aggregation. In the first step, 3.1. Graph representation of users
the users are represented as nodes in a graph structure based on a time-
aware user similarity measure. A deep learning-based community To elucidate the first step in the developed approach, let G = (V, E,
detection system is then developed to identify users who share similar W) be a graph, where V is the node set, E is the edge set indicating the
preferences and belong to the same community. In the third step, the node connections, W is a matrix to indicate the edge weight scores, and
trust network generation leverages the acquired user communities, |V| is the number of all nodes. In this representation, each user is
commonly rated foods, and the ingredients in these foods. In the fourth considered a node in the graph G, where V and E denote the set of all
step, a novel user popularity measure is introduced by employing user users and edges, respectively. Additionally, we construct the weight
centrality measures and trust values between users. This measure aims matrix W using a time-aware similarity criterion, where wu,v ∈ W de­
to determine how to aggregate user preferences within the same group notes the time-aware similarity between users u and v. Considering the
of users. In the next step, a single user rating is predicted by utilizing the changing preferences of users over time is the main benefit of this
health factor of foods and user preference. Finally, in the sixth step, the measure. This measure can be calculated as follows:
preferences of users belonging to the same group are aggregated using ∑
((ru,i − ru ) × (rv,i − rv ) × TW u,v,i )
the calculated user popularity measure. Fig. 1 shows the overall flow
(1)
fi ∈Iu,v
diagram of our suggested HGFR model. In the following subsections, we wu,v = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑ ̅√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑
((ru,i − ru )2 × TW u,v,i ) ((rv,i − rv )2 × TW u,v,i )
describe our developed method in more detail. fi ∈Fu,v fi ∈Fu,v

where Fu,v indicates the food set that was rated by both users u and v; ru,fi
is the rating of food fi allocated by user u; ru indicates the average ratings

4
M. Rostami et al. Neurocomputing 576 (2024) 127326

Fig. 2. Joint learning approach that combines representation learning, autoencoder, and clustering.

of user u; and TWu,v,i denotes the time weight associated with the ratings efficiently learn a compact representation of high-dimensional data by
of users u and v allocated to food fi , which can be calculated as follows: mapping it to a lower-dimensional latent space while preserving the
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ intrinsic structure of the data. This process not only mitigates the issues
TW u,v,i = e− λ(TL− tu,i ) × e− λ(TL− tv,i ) (2) associated with high dimensionality but also helps reveal complex
non-linear relationships that may be obscured in the original space.
where tu,i indicates the timestamp of the rating of food fi allocated by Furthermore, deep learning models can be fine-tuned to be more robust
user u and TL is the maximum value of timestamps in the dataset. To to noise and anomalies, yielding more accurate and coherent clusters.
address the issue of data sparsity, we divided the user ratings into Therefore, for data with a high number of attributes, deep
monthly intervals. Nonetheless, weekly or even daily periods are viable learning-based clustering offers a robust and efficient way to identify
and reliable for abundant user-food ratings. Considering that the data­ patterns and groupings that may be elusive to other methods [62].
sets were collected from 2000 to 2018 (a span of 19 years), the total In studies dealing with data with high dimensions, such as this one,
number of intervals is 228. Moreover, λ is a user specified parameter that dimension reduction [62,63] is typically employed. Deep neural
sets the time factor significance in the time-aware similarity statement. network (DNN) models [64], also referred to as deep clustering models,
High (low) λ values indicate that the time factor has a greater (lower) are regarded as state-of-the-art methods in this field. DNN involves both
influence on similarity values. dimension reduction and clustering, which are sometimes conducted
After generating the graph G = (V, E, W), the developed deep simultaneously as a joint learning scheme for neural network parame­
learning-based clustering technique can be employed to identify users’ ters and clusters, as in [60,65]. Our model follows this scheme, where
communities. representation learning with an autoencoder and clustering is proposed,
as shown in Fig. 2. This figure indicates the application of a loss function
to address the problem of the entire community.
3.2. Deep community detection
Our model has a representation learning phase that aims to train
cluster-friendly latent representations using new data representations
Because clustering is recognized as one of the common approaches
with a stack autoencoder and a cluster estimator step that appraises the
for community detection [60], we developed a clustering algorithm to
optimum number of clusters. Specifically, we use the user similarity
group the nodes (i.e., users) generated in the previous subsection into
matrix generated through Eq. (1) as the starting point for our user vector
distinct communities in this step. For precision, we utilized the user
and then explore new data representations to detect user communities.
similarity matrix, calculated using Eq. (1) as the input of the community
The parameters of these components are jointly optimized in a unified
detection algorithm in this phase. This user similarity matrix was sub­
framework.
jected to our developed community detection algorithm for analysis,
enabling the identification of user communities. The user similarity L = LAut + LClu (3)
matrix from the previous step represents high-dimensional data. Deep
learning-based clustering is suited for such high-dimensional user sim­ where LAut and LClu denote the target measure of the feature learning
ilarity matrix calculus for a variety of compelling reasons [61]. step and estimating cluster step.
High-dimensional data often suffer from the curse of dimensionality, The output of the decoder, which is reconstructed using X ̂ =Ẑ (l) , is
a
which can severely degrade the performance of traditional clustering the representation learning phase. Therefore, the reconstruction loss of
techniques. Deep learning, specifically deep autoencoders, can

5
M. Rostami et al. Neurocomputing 576 (2024) 127326

the autoencoder is calculated as follows: ∑∑ ( )

qik
1 lClu = kl(Q‖P ) = pik log (7)
LAut = ̂ 22
‖X − X‖ (4) i k
P ik
2N
The pseudo-code of the developed user community detection method
Here, X represents the user similarity matrix, which can be calcu­
is highlighted in Algorithm 1.
lated using Eq. (1).
̂ (l) ∈ Rn×d from the initial Algorithm 1. The developed user community detection algorithm.
After calculating the latent representation Z a
user matrix, we can employ it to create the soft cluster assignment

Qa ∈ Rn×k , where Qa is calculated from Eq. (5) and denotes the 3.3. Trust network generation
probability that node vi is allocated to cluster k. Specifically, qaik scores
the similarity between the node representation zai and clustering center The trust relationship is crucial in online social networks because
μak when Student’s t-distribution [66] is used as a kernel. Generally, users can build trust networks that are similar to those in real-life re­
soft assignments are more likely to be trustworthy when nodes are close lationships. Therefore, social interaction between users is based on trust
to the cluster center as follows: statements established from their subjective viewpoints. As people are
( more likely to try food based on recommendations from friends than
⃦ ⃦2 )− 1
1 + ⃦zai − μak ⃦ website advertisements, using information about people’s trust can
qaik = ∑( )− 1 (5) improve the accuracy of recommender systems. Moreover, users’ trust
1 + ‖zai − μak ‖2 statements can be incorporated into trust-based recommender systems,
j
particularly when rating data are sparse for collaborative filtering-based
The distribution of nodes, denoted as Pa , is effectively obtained from models. According to previous studies [67,68], users who trust each
Eq. (6) for enhancing cluster purity by assigning greater weight to a data other tend to display similar rating profiles. Therefore, trust relation­
point that exhibits high confidence. Notably, to attain the initial centers ships can be incorporated into traditional collaborative filtering systems
μak (k = 1; 2; …; K), we initially pre-train the autoencoder by optimizing as supplementary information and insight for predicting user favorites.
the reconstruction loss to achieve a significant representation Zα . Sub­ Due to the absence of a trust network in many food social networks,
sequently, we perform K-means clustering on the learned representa­ trust information should be implicitly extracted from the available data.
tions to initialize the model. However, in subsequent training phases, Additionally, because asymmetric trust information cannot be easily
the K-means clustering is not utilized again, as detailed below: extracted from users’ interactions, this research simply assumes that
/ trust information between two users is symmetric.
∑
q2ij qij In our case, we assume that two users who either share the same food
Paik = /i (6) preference or belong to the same community (as per the previous stage)
∑ 2 ∑
qij qij trust each other. Furthermore, we assume that such a food preference
j i can be captured either by the extent of the common food recipes ranked
by both users or by the total amount of food ingredients contained in the
where q2ij can be calculated from Eq. (5). rated foods by both users. This agrees to some extent with other con­
Our objective is to bring the existing cluster assignments, repre­ sumer psychology studies, which revealed that similar food consump­
sented by Qa , close to the target distribution Pa so that the resulting tion promotes trust and cooperation, such as that by Woolley and
clusters can be calculated using the joint training of autoencoder and Fishbach [69].
cluster assignments. To achieve this, we define the clustering loss based Inspired by this intuition, the following is suggested to quantify trust
on the KL divergence between these two distributions as follows: between user u and user v, where we explicitly account for i) the extent
of common food rated by both individuals, ii) common ingredients in

6
M. Rostami et al. Neurocomputing 576 (2024) 127326

foods rated by both users, and iii) whether both users belong to the same 3.5. User rating prediction
community:
( ) Diet styles and eating habits are the main factors affecting prevent­
1 |Fu,v | |Iu,v |
Trust(u, v) = + + Com(u, v) (8) able non-communicable diseases. The World Health Organization
3 |Fu ∪ Fv | |Iu ∪ Iv |
(WHO) states that non-communicable diseases (NCDs) are typically
chronic conditions that arise from a complex interplay of genetics,
where Fu and Fu,v indicate the set of foods rated by user u and both users
physiology, environment, and behavioral factors [72]. Therefore, health
u and v, respectively. Similarly, Iu and Iu,v denote the set of all in­
factors should be considered in food recommendations alongside user
gredients in the foods rated by user u and both users u and v, respec­
preferences to increase user satisfaction. To address these concerns, we
tively. Com(u, v) is an indicator variable such that Com(u, v) = 1 if users
propose a novel health-aware rate prediction measure that incorporates
u and v are in the same group (community). Otherwise, Com(u, v) = 0.
both health and user preferences to assess the predicted ratings. This
Eq. (8) also suggests that no preference exists between the three key
measure is defined based on two factors: the health factor of foods and
components of our trust measure, which favors arithmetic average
the user’s preference factor. We detail these two factors in the following.
operation in line with the principle of insufficient reason in statistics
Health Factor. This is obtained for a given food by considering its
[70].
seven nutritional facts, namely proteins, carbohydrates, sugars, sodium,
It is evident from Eq. (8) that the introduced trust measure is sym­
fat, saturated fats, and fibers. For each of these nutrition components,
metric and takes values in the unit interval, which eases the interpre­
the World Health Organization (WHO) [70] has provided a range for the
tation of its results.
nutrition amount of a healthy meal. This is summarized in Table 1,
which presents the healthy ranges of each nutritional factor. Therefore,
3.4. Single user popularity calculation the health factor of food fi can be constructed by aggregating the indi­
cator variables of each nutritional factor. Specifically, this is expressed
There is no doubt that popular nodes in communities play a signifi­ as
cant role in social groups. Indeed, popular users play a significant role in
leading the group and guiding its members to a common viewpoint. HF(fi ) = LabelWHO(proteins) + LabelWHO(carbohydrates)
Intuitively, popular nodes are those who are most trusted and have high + LabelWHO(sugars) + LabelWHO(sodium) + LabelWHO(fat)
influence among the group’s users. Therefore, in our developed HGFR + LabelWHO(saturated fats) + LabelWHO(fibers)
model, we consider the trust values and centrality scores of the nodes
(14)
(users) to quantify node popularity in each community. In this study, the
harmonic mean of the two factors—trust score and Node centrality—is where for each nutrition component c, LabelWHO(c) = 1 if the amount
considered to calculate user popularity in the groups as follows: of c falls within the suggested range listed in Table 1; otherwise,
( )
2 × SumTrust(u) × Centrality(u) LabelWHO(c) = 0. Therefore, HF fi ranges between 0 (for fully un­
Popularity(u) = (9) healthy food) and 1 (for fully healthy food).
SumTrust(u) + Centrality(u)
Because the recorded food rates are in the ranges of, [1,5], we should
( )
where SumTrust(u) indicates the total trust value of all group members of also convert HF fi to this range for compatibility. To this end, we
node u. SumTrust value for user u in the group Gc can be computed as normalized the healthy factor of food fi as follows:
follows:
4 × HF(fi )
∑ NHF(fi ) = 1 + (15)
SumTrust(u) = Trust(u, v) (10) 7
Preference Factor. This factor is used to predict user preference
v∈Gc

where, Trust(u, v) is calculated using Eq. (8). with respect to a given food. In our model, users’ communities identified
Centrality(u) denotes the node centrality measure for user u. This by the developed deep learning-based community detection approach
captures the influence of node u in the constructed social network. We are utilized for predicting the preference of each group member. Spe­
used the Laplacian centrality [71] technique to score the centrality of cifically, for each group member, the preferences of users within the
the nodes. For the generated user graph G = (V, E, W) in Section 3.1, X same group are utilized as a proxy to predict the rating of each food.
can be calculated as follows: Formally, let Ku,i be the set of users who belong to the same community
⎛ ⎞ as user u and have provided ratings for food item fi . To some extent, Ku,i
X1 0 ... 0 can be assimilated as the set of (nearest) neighbors of user u with respect
⎜ 0 X2 ... 0 ⎟
X(G) = ⎜ ⎟ (11) to food fi . Therefore, the predicted rating of food fi by user u or,
⎝ . . ⎠
. . equivalently, the preference of user u for food fi can be estimated by
0 0 ... Xn
calculating the marginal contribution across the neighborhood Ku,i by
∑|V| considering the user-to-user similarity scores as follows:
where Xi indicates the sum weight of node vi , calculated as Xi = j=1 wi,j , ∑
where wi,j is the weight of the edge linking nodes i and j computed wu,v × (rv,i − rv )
(16)
v∈Ku,i
using Eq. (1). Furthermore, the Laplacian energy of graph G is defined as Preferenceu,fi = ru + ∑
wu,v
follows: v∈Ku,i

∑|V| ∑
EL (G) = i=1
Xi 2 + 2 i<j w2i,j (12) where Preferenceu,fi is the estimated rating of food fi for user u, ru denotes
the average of ratings recorded by user u, rv,i indicates the rating of food
Finally, the centrality of node u, Centrality(u), can be defined as fi recorded by user v, and wu,v denotes the similarity measure between
follows:
users u and v defined using Eq. (1).
(ΔЕ)ui EL (G) − EL (Gu ) Once the preference rating has been calculated by considering both
Centrality(u) = = (13) the ratings of previous users and the healthy actor associated with a
EL (G) EL (G)
specific food according to its nutrition facts, the final rating of food fi for
where Gu is the graph created by removing u from graph G. user u is computed as a convex combination of health factor and pref­
Finally, (13) and (10) are substituted in (9) to calculate erence factor as follows:
Popularity(u)..

7
M. Rostami et al. Neurocomputing 576 (2024) 127326

Pu,fi = (γ).NHF(fi ) + (1 − γ).Preferenceu,fi . (17) aggregating the group rating over all foods items, the top k foods that
have the maximum rates will be suggested to the group.
( )
where HF fi is the health factor of food fi and .Prefrenceu,i is the pref­ The overall steps of the suggested HGFR model are shown in Algo­
erence factor of user u for food fi that can be calculated using Eqs. (15) rithm 2.
and (16), respectively. The parameter γ balances between the user Algorithm 2. Healthy Group Food Recommendation (HGFR).
health factor and preference factor and takes values in the unit interval.

This enables us to adjust the recommendation to accommodate the 4. Experiments

user’s view on the balance between the health factor and the user’s
preference according to the value assigned to weight γ. In other words, In this section, we examine the developed HGFR by conducting a
by adjusting γ, group members can decide on the tradeoff between fa­ wide range of experiments on two real-food social network datasets to
vorites and the health measure in the final food recommendation system compare its performance with that of other food recommendation
such that small (resp. high) values of γ will favor the user’s preference models. During the experiments, we compared the proposed HGFR with
(resp. health factor). In the developed controllable food recommenda­ the following system for group recommendation:
tion model, users can be involved in the recommendation process by
modifying the contribution of the health factor parameter. • GRGP [55]: A group recommendation model based on genre favor­
ites that reduces clustering algorithm complexity.
• ECOagg [56]: It predicts the preferences of groups based on back­
3.6. User group rating aggregation ground knowledge about group members.

After predicting unknown ratings for all members of each group

(community), the final group ratings must be aggregated as well.
Accordingly, for each food, the developed system utilizes a simple
weighted averaging mechanism of the prediction scores based on the Table 1
user popularity scores. Specifically, the aggregated rating of group Gc for Nutrition components and their suggested ranges [72].
food fi is calculated as follows: Dietary factor Suggested range
∑( ) Proteins 10-15%
Pu,fi × Popularity(u)
Carbohydrates 55-75%
(18)
u∈Gc
GPGc ,fi = ∑ Sugars < 10%
Popularity(u)
u∈Gc Sodium <5g
Fat 15-30%
where Pu,fi indicates the health-aware single user rating of user u for food Saturated fats < 10%
Fibers > 3%
fi calculated using (17) and |Gc | is the number of users in group Gc . After

8
M. Rostami et al. Neurocomputing 576 (2024) 127326

Table 2 described in Table 2.

Details of datasets from Allrecipes.com and Food.com.
Characteristics Allrecipes.com Food.com 4.2. Evaluation metrics
#Users 68,768 25,076
#Foods 45,630 178,265 For assessing the recommendation performance of compared sys­
#Ratings 1,093,845 749,053 tems, four single user-based evaluation metrics, including precision,
#Ingredients 5073 33,147 recall, normalized discounted cumulative gain (NDCG), and health, as
well as one group satisfaction metric, are used.
Precision, recall, and NDCG are commonly used metrics for
• DP_MCF [57]: A group recommender system for aggregating deep
measuring the efficiency of RSs. Precision and recall metrics are calcu­
profiles of group members instead of item preference-based
lated as the ratio of relevant recommendations as follows:
recommendations.
• HVGR [58]: A hierarchy visualization-based group recommendation |{Recommended items that are relevant}|
Precision = (19)
for providing more information than simple recommendations using L
multilevel nodes and edges
• CAGR [59]: A group recommendation model to alleviate the vital Recall =
|{Recommended items that are relevant}|
(20)
group and user-item data sparsity issues |{all relevant items}|

where L denotes the length of the recommendation list (i.e., how many
foods are recommended).
4.1. Datasets NDCG assesses the capability of the recommendation system to
generate a list of suggestions featuring appropriate items at the top:
A comparison of the group recommender systems is conducted using DCG
two datasets obtained from https://www.allrecipes.com and htt NDCG = (21)
DCGmax
ps://www.food.com/ [3,15]. In the Allrecipes.com dataset, 52,821
recipes were extracted. Moreover, users’ ratings, food nutrition infor­ where,
mation, and timestamps for ratings were collected for each recipe. A
∑
L
preprocessing operation was performed on this dataset to obtain infor­ DCG = rel1 +
reli
(22)
mation related to 68,768 users, 45,630 foods, and 1,093,845 ratings. i=2
log2 (i + 1)
Data obtained from Food.com includes 25,076 users, 178,265 recipes,
and 749,053 ratings. Using the Allrecipes dataset, seven nutrition and reli is used to specify whether the suggested item fi is relevant.
components were analyzed: proteins, carbohydrates, sugars, sodium, Therefore, if the suggested food fi is relevant, then reli = 1; otherwise,
fat, saturated fats, and fibers. Due to the insufficient knowledge about reli = 0. Moreover, DCGmax can be calculated as follows:
the fiber nutrition factor in the Food.com dataset, we assume this ∑
L
1
component is within an ideal range for all foods in this dataset. In both DCGmax = 1 + (23)
log2 (i + 1)
datasets, users assigned a rating of 1 to 5 to the foods. These datasets are i=2

Fig. 3. Evaluation of the proposed HGFR model with different λ parameter values and the length of recommendations list L = 10 for Allrecipes.com and Food.
com datasets.

9
M. Rostami et al. Neurocomputing 576 (2024) 127326

Fig. 4. Evaluation of the developed HGFR model with different γ parameter values and the length of recommendations list L = 10 for Allrecipes.com and Food.
com datasets.

In addition to the aforementioned accuracy criterion, a new criterion group c and the length of the recommendation list, respectively. The
derived from health levels of the recommended foods is employed to value of GSM is in the range [0,1], where GSM = 0 indicates that none of
measure the healthiness of the recommended foods. The health factor of the group members is satisfied with the recommended foods, whereas
the suggested list is evaluated as follows: GSM = 1 indicates all of the group members are satisfied.
∑
L
NHF(fi )
Health = i=1 (24) 4.3. Parameter setting and analysis
L
( ) Tuning the input parameters is one of the most important aspects of
where NHF fi denotes the health score of food fi , which can be defined
evaluating an algorithm. As part of this section, we conduct a sensitivity
using Eq. (15). Notably, the value of the health measure is in the range [0, analysis to determine the best values for the input parameters of the
1], where Health = 0 means that the recommended food list is entirely developed HGFR model.
unhealthy, whereas Health = 1 denotes that the recommended foods are There are two input parameters in the developed HGFR model: (1) λ
fully healthy. is a time weight parameter used in Eq. (2) to control the impact of the
Furthermore, a group satisfaction metric (GSM) is computed to time factor in the time-aware similarity measure calculation, and (2) γ is
measure the recommendation list suggested to the groups based on each a user-defined parameter in Eq. (17) to balance between the user’s
member’s level of satisfaction. The GSM measure can be defined as preference and food health terms.
follows: Fig. 3 illustrates the sensitivity analysis of the λ parameter, where its
∑
(FR ∩ SF u ) value ranges from 0 to 4 with a step size of 0.5. Notably, that λ = 0
GSM(GC ) =
u∈Gc
(25) means that the time factor will be ignored in the calculus of the simi­
( |GC | × L ) larity measure. Fig. 3 shows that the time factor positively affects HGFR
performance, as all evaluation criteria increased when λ increased.
where FR is the set of recommended foods and SF u is the list of foods that Therefore, it can be concluded that incorporating the changes in diet
received ratings of 4 or 5. Moreover, Gc and L indicate the set of users in over time into the recommendation process results in a more accurate

Table 3
Evaluation of our suggested HGFR compared to other recommendation systems on the Allrecipes.com dataset in terms of single user-based evaluation metrics.
Methods Precision Recall NDCG

L= 10 L= 15 L= 20 L= 10 L= 15 L= 20 L= 10 L= 15 L= 20

GRGP 0.0582 0.0452 0.0323 0.0601 0.0691 0.0781 0.0382 0.0434 0.0459
ECOagg 0.0693 0.0515 0.0412 0.0674 0.0854 0.1022 0.0456 0.0489 0.0513
DP_MCF 0.0672 0.0569 0.0491 0.0639 0.0882 0.1059 0.0432 0.0548 0.0587
HVGR 0.0711 0.0581 0.0532 0.0681 0.0893 0.1106 0.0469 0.0604 0.0718
CAGR 0.0720 0.0638 0.0556 0.0741 0.0927 0.1177 0.0491 0.0622 0.0742
HGFR 0.0802 0.0792 0.0723 0.0863 0.1050 0.1292 0.0578 0.0703 0.0794

10
M. Rostami et al. Neurocomputing 576 (2024) 127326

Table 4
Evaluation of our suggested HGFR compared to other recommendation systems on the Food.com dataset in terms of single user-based evaluation metrics.
Methods Precision Recall NDCG

L= 10 L= 15 L= 20 L= 10 L= 15 L= 20 L= 10 L= 15 L= 20

GRGP 0.0492 0.0369 0.0263 0.0513 0.0581 0.0658 0.0312 0.0352 0.0378
ECOagg 0.0501 0.0403 0.0281 0.0527 0.0637 0.0761 0.0341 0.0361 0.0415
DP_MCF 0.0545 0.0438 0.0320 0.0559 0.0681 0.0754 0.0368 0.0387 0.0416
HVGR 0.0629 0.0508 0.0441 0.0653 0.0809 0.1021 0.0442 0.0541 0.0624
CAGR 0.0661 0.0571 0.0468 0.0682 0.0860 0.1071 0.0471 0.0568 0.0688
HGFR 0.0728 0.0684 0.0594 0.0767 0.0968 0.1132 0.0562 0.0665 0.0729

Table 5 Table 6
Evaluation of our HGFR model compared to other models on the Allrecipes.com Evaluation of different models on the Allrecipes.com and Food.com datasets in
and Food.com datasets in terms of health measure. terms of GSM measure.
Methods Allrecipes.com Food.com Group Size Methods Allrecipes.com Food.com

L= 10 L= 15 L= 20 L= 10 L= 15 L= 20 L= 10 L= 15 L= 20 L= 10 L= 15 L= 20

GRGP 0.3821 0.3568 0.3341 0.4005 0.3687 0.3324 |GC |¼5 GRGP 84.57 79.67 73.57 83.17 82.14 78.94
ECOagg 0.3954 0.4005 0.3844 0.4184 0.3905 0.3238 ECOagg 87.24 82.68 80.06 84.65 83.07 77.64
DP_MCF 0.3478 0.3205 0.2924 0.3569 0.3324 0.2984 DP_MCF 86.68 81.49 78.68 85.99 82.67 80.08
HVGR 0.3135 0.2968 0.2636 0.3218 0.2815 0.2687 HVGR 89.01 84.67 79.67 86.61 83.48 81.97
CAGR 0.3006 0.2754 0.2425 0.2954 0.2705 0.2327 CAGR 89.87 85.37 82.48 88.36 85.19 82.36
HGFR 0.5103 0.4812 0.4557 0.4732 0.4438 0.3963 HGFR 94.65 90.09 87.19 91.32 88.36 85.49
|GC |¼10 GRGP 81.78 79.94 74.94 80.98 76.55 72.34
ECOagg 84.59 80.08 76.07 82.66 81.60 78.36
food recommendation. In our experiments, however, the performance DP_MCF 85.68 80.48 75.36 83.58 81.48 77.29
degrades if λ is greater than 2.5 and 3 for Allrecipes.com and Food.com, HVGR 86.57 81.46 76.38 84.32 80.61 78.98
CAGR 87.63 82.27 78.68 85.78 82.27 79.26
respectively. This is primarily because the recommendations are done HGFR 90.24 86.79 82.48 89.31 86.76 82.69
only based on the recent ratings of users, and most of their historical user |GC |¼20 GRGP 72.92 71.35 69.68 70.19 69.67 67.73
ratings are ignored by the model when λ exceeds these values. Therefore, ECOagg 74.02 72.61 70.08 72.44 70.81 68.16
maintaining the HGFR model at its highest level of efficiency by DP_MCF 74.68 71.84 70.54 71.97 69.49 68.91
HVGR 75.68 73.22 71.36 73.38 72.36 69.04
balancing this parameter is crucial. As part of the experiments, λ = 2.5
CAGR 76.68 74.06 71.82 74.18 74.49 70.08
and λ = 3 were initialized for the Allrecipes.com and Food.com datasets, HGFR 81.38 78.69 75.81 79.68 77.48 74.96
respectively, to evaluate the performance of the developed HGFR model.
The impact of the healthy parameter of γ is investigated in the second
part of the sensitivity analysis. In the final recommendations, healthy HGFR has better performance than the second-best model by approxi­
foods are more important when this parameter is close to 1. When the mately 5.94%, 12.46%, and 19.32% for L = 10 based on the single user-
health parameter is set to zero, our model only considers the user’s fa­ based evaluation metrics of precision, recall, and NDCG, respectively,
vorites. In general, increasing the impact of the health parameter on the for the Food.com dataset. In accordance with all utilized single user-
recommendations can lead to a decrease in the system efficiency. This is based evaluation metrics and datasets, our developed HGFR model
because users tend to try and review unhealthy recipes more frequently. will significantly outperform other recommendation models, as illus­
Consequently, suggesting healthy recipes to such users can adversely trated in Tables 5 and 6.
affect the overall efficiency of the model. Fig. 4 denotes the sensitivity Next, a comparison based on health factor is conducted, which is
analysis of γ parameter on the various performance (i.e., precision, another single user-based evaluation criterion. The results are listed in
recall, and NDCG) and health measures. The results indicated that Table 5 for Allrecipes.com and Food.com. The results in Table 5 suggest
increasing the value of the healthy parameter from zero to one reduces that our developed recommendation system is the best model, as health
the precision, recall, and NDCG of HGFR by 27.99%, 25.03%, and levels of 0.5103, 0.4812, and 0.4557 were achieved for the Allrecipes.
36.07%, respectively, and increases the health factor by 40.63%. com dataset in the recommendation length of 10, 15, and 20, respec­
tively. In this case, the ECOagg recommendation system had the second-
best health levels of 0.3954, 0.4005, and 0.3844 for the Allrecipes.com
4.4. Performance comparison
dataset in the recommendation length of 10,15, and 20, respectively.
Therefore, our developed model achieved performance improvements of
This section compares the HGFR method with other recommendation
29.05%, 20.14%, and 18.54% for L = 10, L = 15, and L = 20 versus the
approaches using a variety of evaluation metrics and the different
second-best performer (i.e., ECOagg) on the Allrecipes.com dataset.
lengths of the recommendations list (L). Initially, the accuracy-based
Moreover, Table 5 reports that the HGFR performed better than other
single user-based evaluation metrics, namely precision, recall, and
systems in the Food.com dataset, with performance improvements of
NDCG, are used to evaluate the accuracy measures of the food recom­
approximately 13.09%, 13.64%, and 22.39% for the recommendation
mendations generated by the compared systems. Tables 3 and 4 sum­
length of 10, 15, and 20 compared to the next best system (i.e., ECOagg).
marize the experimental results for Allrecipes.com and Food.com datasets
We can conclude from this part of the experiments and based on single
for these metrics. The HGFR system achieved the highest precision,
user-based evaluation metrics including precision, recall, NDCG, and
recall, and NDCG measures, indicating that it is the best performer. The
health that the suggested HGRF not only significantly improves food
results in Table 4 and for the recommendation list of L = 10 reveal that
recommendation accuracy but also makes healthier recommendations
HGFR achieves the values of 0.0802, 0.0863, and 0.0578 for the preci­
than other models.
sion, recall, and NDCG measures, respectively. For the same case, CAGR
The precision, recall, NDCG, and health measures that have been
(i.e., the second-best model) yields scores 0.0720, 0.0741, and 0.0491,
utilized for the evaluation above are employed to assess the single user-
respectively. Compared to the second-best model, the developed system
based performance metrics. In the next experiment, the GSM measure is
exhibits improvements of 11.38%, 13.76%, and 17.71% for precision,
used to assess the degree of satisfaction of the groups. Table 6 indicates
recall, and NDCG, respectively. Moreover, Table 4 demonstrates that

11
M. Rostami et al. Neurocomputing 576 (2024) 127326

Fig. 5. A comparison between the developed HGFR and different scenarios based on Precision, Recall, NDCG, and GSM measures.

the satisfaction degree for each group with a group size of 5, 10, and 20 Scenario 2: In this scenario, it is assumed that the developed deep
and a recommendation length of 10,15, and 20, respectively. Table 6 learning-based user community detection model is ignored in the
confirms that the developed HGFR outperforms other methods on all recommendation process. In this case, the classic K-nearest neighbor
efficiency criteria and recommendation lengths. Based on comparisons approach is used instead of the developed community detection model
with other models, the HGAT system has the second-best performance. to predict the final single-user ratings (Eq. (16)). This scenario can be
The superiority of HGFR over HGAT indicates the considerable capa­ used to evaluate the effectiveness of the proposed user community
bility of the developed model to produce a higher group satisfaction detection model in improving recommendation performance.
metric (GSM). The superiority of the suggested system over other sys­ Scenario 3: This scenario aims to determine the importance of the
tems is primarily attributable to the use of a deep learning-based com­ developed user popularity-based measure for aggregating users’ pref­
munity detection approach, which leads to the removal of inefficient erences in groups. To this end, we ignore this popularity measure in the
neighbors from the neighborhood. Furthermore, we calculated the final preference aggregation process. Specifically, in this scenario, the
group aggregation using a novel user popularity measure, which aggregated rating of the group is calculated based on the average user
improved final group satisfaction. rating of the group instead of considering the developed trust-aware user
The benefits of the suggested user group aggregation approach in popularity measures employed in the user rating aggregation (Eq. (18)).
boosting food recommender services to make more accurate and healthy Scenario 4: This scenario aims to determine the importance of the
recommendations is evident. Moreover, the proposed deep learning- developed trust network generation for calculating users’ popularity in
based community detection plays a significant role in identifying the groups. To this end, we ignore the trust measure in the popularity
neighbors of target users in the single-user rating prediction phase. calculation process. Specifically, in this scenario, the popularity of each
member in the group will only be computed based on the user centrality
4.5. Ablation study instead of integrating the trust measure and centrality measure in the
user popularity calculation in the group (Eq. (8)).
Our developed food recommendations makes three main contribu­ According to the above-mentioned scenarios, some experiments
tions: a rating profile expansion technique, a user community detection were designed on the Allrecipes.com and Food.com datasets in terms of
model, and a health-aware reliability measure. A significant part of the different assessment measurements. Fig. 5 illustrates the results of the
analysis is measuring the importance of each contribution in enhancing ablation study performed for these two datasets with a recommendation
the quality of food recommendations. This section performs an ablation length of 10 and a group size of 5. According to these results, Scenario 3
study to demonstrate how the performance of the suggested system has the lowest performance, which is a result of ignoring the proposed
changes when each contribution is ignored. As a result, the following aggregation of users’ preferences into groups during the group recom­
three scenarios are defined: mendation process. Accordingly, we can conclude that the developed
Scenario 1: The time-aware user similarity calculation is ignored in aggregation approach has a critical role to play in the proposed food
this scenario. In other words, The time of all voted user ratings is recommendation system and has a significant effect on the final rec­
considered the same in the user graph representation phase (Eq. (1)). ommendation’s performance. As a result of not incorporating the
Accordingly, this scenario can be useful in identifying users’ potential developed approach for aggregating the preferences of users in the
changes in food preferences, diet, or lifestyle over time and, thus, group, the traditional average preference aggregation technique is used
determining the effect of the time-aware model on the performance of rather than the developed user popularity-based technique. However,
the developed recommendation system. since the developed aggregation of users’ preferences impacts the group

12
M. Rostami et al. Neurocomputing 576 (2024) 127326

satisfaction metric more directly, the improvement rate for the GSM 4.7. Implications
measure is higher than that of the single-user measures (i.e., precision,
recall, and NDCG). Food recommendation systems have gained popularity in recent
Moreover, the developed method based on Scenario 2 performs years due to the increasing demand for personalized nutrition and the
worse than that based on Scenario 1 and Scenario 4 on all single-user need for easy access to healthy food options. These systems utilize al­
measures as well as the group satisfaction metric. This implies that gorithms that provide users with personalized recommendations based
deep learning-based community detection is more important than time- on their dietary preferences, health goals, and restrictions.
aware user similarity calculation and trust-aware popularity measure­ The World Health Organization suggests promoting healthy eating
ment in enhancing the effectiveness of the recommender system. and decreasing preventable noncommunicable diseases by creating
Therefore, when we remove community detection from the recom­ nutritional profiling systems that enable healthier choices. This can only
mendation process, the developed system performs much worse than if be achieved by developing healthy and personalized recommendations.
we ignore the time-aware user graph representation or trust-aware user These systems must also be adaptable to different individual circum­
popularity calculation. As far as all datasets are concerned, HGFR per­ stances and conditions. For this purpose, accounting for health factors in
forms better in all four scenarios for both single-user measures and the the user’s food preference as well as individual’s health condition is
group satisfaction metric. We can therefore conclude that the quality of paramount. Our developed healthy food recommender system provides
recommendations improves when we consider all components of the both a practical and conceptual framework to balance the user’s food
developed system together. preference and the health factor of the foods simultaneously.
The theoretical implications of these systems are broad and include a
4.6. Discussion better understanding of the complex relationship between food and
health and how personalized recommendations can positively impact
The results of our experiments on real social networks and practical individual’s eating habits. On a practical level, these healthy food
recommendation datasets demonstrate that the HGFR model is highly recommendation systems can help individuals make informed decisions
effective in predicting the preferences of both individuals and groups about their food choices, leading to improved dietary patterns and better
and recommending items to them. Our model’s performance, both for health outcomes. By promoting healthy eating habits and providing
individual and group measures, was significantly better than that of the tailored food recommendations, these systems can decrease the risk of
most recent food recommendation models, as listed in Tables 5 and 6. chronic diseases, enhance well-being, and ultimately promote a healthy
Additionally, our developed recommendation system was the top- lifestyle in general. This could potentially help advance the United Na­
performing model in terms of health factors for both datasets tions’ Sustainable Development Goal 3, which targets increasing health
(Table 5) because it is time-aware and considers user communities and and well-being.
changes in a user’s preferences and eating habits over time, unlike the
compared systems, which overlook the time factor in historical user 5. Conclusion
ratings. Therefore, our model accurately reflects real-life scenarios.
Moreover, an ablation experiment was designed to assess the effective­ Online food services have recently harnessed food recommendation
ness of each of our four important novelties (i.e., time-aware user sim­ systems to provide users with customized food recommendations.
ilarity calculation, deep learning-based user community detection, user Despite this, there are still two main challenges to overcome in this area.
popularity-based calculation, and trust network generation) available in The first one is related to group food recommendations. According to our
HGFR. The results of the ablation study on both datasets are presented in knowledge, all traditional food recommendation models have been
Fig. 5, which indicates that the user popularity-based measure and deep designed for single-user scenarios. On the other hand, one of the primary
learning-based user community detection developed for final group obstacles hindering the widespread adoption of food recommender
recommendation are highly effective in improving both the single and systems is the limitation related to recommending healthy food options.
group criteria in the recommender systems. Because nutrition information has not been incorporated into many
Although our proposed model outperformed several state-of-the-art previous food recommendation systems, the recommendations cannot
food recommendation algorithms, there is still room for further guarantee that the foods suggested are healthy for users. To address
improvement, as discussed below: these challenges, this study developed an innovative healthy group food
recommendation model to consider user popularity and the health of
• Although nutrition and diet are vital to people’s health, food foods in a group recommendation mechanism. The main novelty of this
recommendation models are complex due to the diverse range of study is the measure developed to calculate the user’s popularity in the
users. Consequently, fairness, bias, and ethical considerations may group, which is defined based on both trust networks and user centrality.
require more attention in future research. Food recommendation This measure is used to aggregate the preferences of different users in
models have grown increasingly important in people’s lives and, the group. Furthermore, a time-aware similarity measure is defined and
therefore, public health, thus making it imperative to address these then utilized in a deep community detection, which is employed in a
concerns in future research, as they have not been adequately single-user rating prediction. The developed system is measured through
addressed in current models. different experiments and compared with other novel group recom­
• To achieve personalized and balanced food recommendations, it is mendation models to demonstrate its effectiveness. The evaluation
necessary to balance food health and preference measures and pro­ considers various metrics, including accuracy, health factors, and user
vide users with explanations for why a food was suggested or why satisfaction, for both single-user and group scenarios. The results
others were not suggested. Explainable healthy recipe recommen­ revealed that our suggested system outperformed other systems in all
dation systems can be explored in future studies. aspects. Moreover, an ablation study is designed to assess the individual
• Most previous models, including our developed system, do not contributions of the proposed components. The study reveals that pro­
consider user characteristics when recommending foods. Incorpo­ posed contributions have a significant impact on enhancing the perfor­
rating additional user information will be explored in future mance of the food recommendation system.
research.
• Overemphasizing recommendation accuracy may lead to inappro­ CRediT authorship contribution statement
priate recommendations and overfitting. Future research will focus
on utilizing novelty and diversity to address the overfitting issue and Mehrdad Rostami: Conceptualization, Methodology, Software,
enhance the user’s experience. Formal analysis, Writing – original draft, Writing – review & editing.

13
M. Rostami et al. Neurocomputing 576 (2024) 127326

Kamal Berahmand: Methodology, Writing – original draft, Writing – [23] X. Gao, et al., Food recommendation with graph convolutional network, Inf. Sci.
584 (2022) 170–183.
review & editing. Saman Forouzandeh: Methodology, Writing – orig­
[24] T.N.T. Tran, et al., An overview of recommender systems in the healthy food
inal draft, Writing – review & editing. Sajad Ahmadian: Methodology, domain, J. Intell. Inf. Syst. 50 (2018) 501–526.
Writing – original draft, Writing – review & editing. Vahid Farrahi: [25] C. Trattner, D. Elsweiler, Investigating the healthiness of internet-sourced recipes:
Writing – original draft, Writing – review & editing, Supervision. implications for meal planning and recommender systems, Proc. 26th Int. Conf.
World Wide Web (2017).
Mourad Oussalah: Writing – original draft, Writing – review & editing, [26] F. Rehman, et al., Diet-right: a smart food recommendation system, KSII Trans.
Supervision. Internet Inf. Syst. (TIIS) 11 (6) (2017) 2910–2925.
[27] R. Maia, J.C. Ferreira, Context-aware food recommendation system, Int. Assoc.
Eng. 5 (2018) 349–356.
Declaration of Competing Interest [28] R. Sookrah, J.D. Dhowtal, S.D. Nagowah, A DASH diet recommendation system for
hypertensive patients using machine learning. 7th International Conference on
Information and Communication Technology (ICoICT), IEEE, 2019.
The authors declare that they have no known competing financial [29] M. Rostami, et al., Gene selection for microarray data classification via multi-
interests or personal relationships that could have appeared to influence objective graph theoretic-based method, Artif. Intell. Med. 123 (2022) 102228.
the work reported in this paper. [30] S. Azadifar, et al., Graph-based relevancy-redundancy gene selection method for
cancer diagnosis, Comput. Biol. Med. 147 (2022) 105766.
[31] S. Wang, et al., Detecting shilling groups in online recommender systems based on
Data availability graph convolutional network, Inf. Process. Manag. 59 (5) (2022) 103031.
[32] M. Hosseinpoor, et al., Proposing a novel community detection approach to
identify cointeracting genomic regions, Math. Biosci. Eng. 17 (3) (2020)
Data will be made available on request. 2193–2217.
[33] M. Rostami, K. Berahmand, S. Forouzandeh, A novel community detection based
Acknowledgments genetic algorithm for feature selection, J. Big Data 8 (1) (2021) 2.
[34] M. Rostami, M. Oussalah, A novel attributed community detection by integration
of feature weighting and node centrality, Online Soc. Netw. Media 30 (2022)
The project is supported by the Research Council of Finland (former 100219.
Academy of Finland) and Profi5 DigiHealth Research Program (project [35] L. Paleti, P. Radha Krishna, J.V.R. Murthy, Approaching the cold-start problem
using community detection based alternating least square factorization in
number 326291). recommendation systems, Evolut. Intell. 14 (2) (2021) 835–849.
[36] I. Viktoratos, A. Tsadiras, N. Bassiliades, Combining community-based knowledge
References with association rule mining to alleviate the cold start problem in context-aware
recommender systems, Expert Syst. Appl. 101 (2018) 78–90.
[37] C. Chen, C. Toumazou, Personalized expert recommendation systems for optimized
[1] M. Ge, et al., Using tags and latent factors in a food recommender system.
nutrition, Trends Pers. Nutr. 46 (2019) 309–338.
Proceedings of the 5th International Conference on Digital Health, ACM, 2015.
[38] B. Deebak, F. Al-Turjman, A novel community-based trust aware recommender
[2] J. Freyne, S. Berkovsky, Recommending food: Reasoning on recipes and
systems for big data cloud service networks, Sustain. Cities Soc. 61 (2020) 102274.
ingredients. International Conference on User Modeling, Adaptation, and
[39] K. Wang, et al., E-commerce personalized recommendation analysis by deeply-
Personalization, Springer, 2010.
learned clustering, J. Vis. Commun. Image Represent. 71 (2020) 102735.
[3] C. Lin, T. Kuo, S. Lin, A content-based matrix factorization model for recipe
[40] F. Zhang, et al., Graph embedding-based approach for detecting group shilling
recommendation. Pacific-Asia Conference on Knowledge Discovery and Data
attacks in collaborative recommender systems, Knowl. Based Syst. 199 (2020)
Mining, Springer, 2014.
105984.
[4] A. Al-Nazer, T. Helmy, M. Al-Mulhem, User’s profile ontology-based semantic
[41] Y. Leng, L. Yu, Incorporating global and local social networks for group
framework for personalized food and nutrition recommendation, Procedia Comput.
recommendations, Pattern Recognit. 127 (2022) 108601.
Sci. 32 (2014) 101–108.
[42] S. Ding, et al., Time-aware cloud service recommendation using similarity-
[5] D. Bianchini, et al., PREFer: a prescription-based food recommender system,
enhanced collaborative filtering and ARIMA model, Decis. Support Syst. 107
Comput. Stand. Interfaces 54 (2017) 64–75.
(2018) 103–115.
[6] M. Ge, F. Ricci, D. Massimo, Health-aware food recommender system. Proceedings
[43] P. Kefalas, Y. Manolopoulos, A time-aware spatio-textual recommender system,
of the 9th ACM Conference on Recommender Systems, ACM, 2015.
Expert Syst. Appl. 78 (2017) 396–406.
[7] M. Izadi, M. Nejati, A. Heydarnoori, Semantically-enhanced topic recommendation
[44] F. Rezaeimehr, et al., TCARS: time-and community-aware recommendation system,
systems for software projects, Empir. Softw. Eng. 28 (2) (2023) 50.
Future Gener. Comput. Syst. 78 (2018) 419–429.
[8] Z. Lei, et al., Is the suggested food your desired?: Multi-modal recipe
[45] Z. Cui, et al., Personalized recommendation system based on collaborative filtering
recommendation with demand-based knowledge graph, Expert Syst. Appl. 186
for IoT scenarios, IEEE Trans. Serv. Comput. 13 (4) (2020) 685–695.
(2021) 115708.
[46] D. Sánchez-Moreno, Y. Zheng, M.N. Moreno-García, Time-aware music
[9] N. Jia, J. Chen, R. Wang, An attention-based convolutional neural network for
recommender systems: modeling the evolution of implicit user preferences and
recipe recommendation, Expert Syst. Appl. 201 (2022) 116979.
user listening habits in a collaborative filtering approach, Appl. Sci. 10 (15) (2020)
[10] M.S.J. Rokon, et al., Food recipe recommendation based on ingredients detection
5324.
using deep learning, Int. Conf. Comput. Adv. (ICCA 2022) (2022).
[47] J. Zhao, et al., TBTF: an effective time-varying bias tensor factorization algorithm
[11] S. Shabanabegum, et al., IOT enabled food recommender with NIR system, Mater.
for recommender system, Appl. Intell. 51 (7) (2021) 4933–4944.
Today.: Proc. 1 (2020) 1–10.
[48] A.N. Ngaffo, W. El Ayeb, Z. Choukair, A time-aware service recommendation based
[12] Y. Pérez-Almaguer, R. Yera, A.A. Alzahrani, L. Martínez, Content-based group
on implicit trust relationships and enhanced user similarities, J. Ambient Intell.
recommender systems: a general taxonomy and further improvements, Expert Syst.
Humaniz. Comput. 12 (2) (2021) 3017–3035.
Appl. 184 (2021) 115444.
[49] S. Ahmadian, et al., Alleviating data sparsity problem in time-aware recommender
[13] R. Beaglehole, et al., Priority actions for the non-communicable disease crisis,
systems using a reliable rating profile enrichment approach, Expert Syst. Appl. 187
Lancet 377 (9775) (2011) 1438–1447.
(2022) 115849.
[14] M. Rostami, M. Oussalah, V. Farrahi, A novel time-aware food recommender-
[50] S. Feng, et al., Relation-aware dynamic attributed graph attention network for
system based on deep learning and graph clustering, IEEE Access (2022).
stocks recommendation, Pattern Recognit. 121 (2022) 108119.
[15] X. Gao, et al., Food recommendation with graph convolutional network, Inf. Sci.
[51] R. Yera, A.A. Alzahrani, L. Martínez, A fuzzy content-based group recommender
584 (2022) 170–183.
system with dynamic selection of the aggregation functions, Int. J. Approx. Reason.
[16] X. Gao, et al., Hierarchical attention network for visually-aware food
150 (2022) 273–296.
recommendation, IEEE Trans. Multimed. 22 (6) (2020) 1647–1659.
[52] M. Riyahi, M.K. Sohrabi, Providing effective recommendations in discussion groups
[17] T. Osadchiy, I. Poliakov, P. Olivier, M. Rowland, E. Foster, Recommender system
using a new hybrid recommender system based on implicit ratings and semantic
based on pairwise association rules, Expert Systems with Applications 115 (2019)
similarity, Electron. Commer. Res. Appl. 40 (2020) 100938.
535–542.
[53] L. Baltrunas, T. Makcinskas, F. Ricci, Group recommendations with rank
[18] L. Meng, et al., Heterogeneous fusion of semantic and collaborative information for
aggregation and collaborative filtering, Proc. Fourth ACM Conf. Recomm. Syst.
visually-aware food recommendation, Proc. 28th ACM Int. Conf. Multimed. Seattle
(2010).
WA, USA (2020).
[54] E. Yalcin, F. Ismailoglu, A. Bilge, An entropy empowered hybridized aggregation
[19] S. Ahmadian, et al., Healthy food recommendation using a time-aware community
technique for group recommender systems, Expert Syst. Appl. 166 (2021) 114111.
detection approach and reliability measurement, Int. J. Comput. Intell. Syst. 15 (1)
[55] Y.-D. Seo, et al., Group recommender system based on genre preference focusing on
(2022) 1–24.
reducing the clustering cost, Expert Syst. Appl. 183 (2021) 115396.
[20] R.Y. Toledo, A.A. Alzahrani, L. Martinez, A food recommender system considering
[56] F. Ismailoglu, Aggregating user preferences in group recommender systems: a
nutritional information and user preferences, IEEE Access 7 (2019) 96695–96711.
crowdsourcing approach, Decis. Support Syst. 152 (2022) 113663.
[21] L. Yang, et al., Yum-me: a personalized nutrient-based meal recommender system,
[57] L.N.H. Nam, Profile aggregation-based group recommender systems: moving from
ACM Trans. Inf. Syst. (TOIS) 36 (1) (2017) 1–31.
item preference profiles to deep profiles, IEEE Access 10 (2022) 6218–6245.
[22] E. Asani, H. Vahdat-Nejad, J. Sadri, Restaurant recommender system based on
sentiment analysis, Mach. Learn. Appl. 6 (2021) 100114.

14
M. Rostami et al. Neurocomputing 576 (2024) 127326

[58] W. Wang, G. Zhang, J. Lu, Hierarchy visualization for group recommender systems, Saman Forouzandeh is currently pursuing a Ph.D. degree at
IEEE Trans. Syst. Man Cybern. Syst. 49 (6) (2019) 1152–1163. the School of Mathematics and Statistics, University of New
[59] H. Yin, et al., Overcoming data sparsity in group recommendation, IEEE Trans. South Wales (UNSW) in Sydney, NSW, Australia. His primary
Knowl. Data Eng. 34 (7) (2022) 3447–3460. research interests lie in machine learning, graph neural net­
[60] S. Molaei, et al., Deep node clustering based on mutual information maximization, works, and recommender systems. In 2013, he obtained a
Neurocomputing 455 (2021) 274–282. master’s degree in software engineering from the Islamic Azad
[61] E. Eskandarnia, H.M. Al-Ammal, R. Ksantini, An embedded deep-clustering-based University, Science and Research Branch, located in Tehran,
load profiling framework, Sustain. Cities Soc. 78 (2022) 103618. Iran.
[62] X. Peng, et al., Deep clustering with sample-assignment invariance prior, IEEE
Trans. Neural Netw. Learn. Syst. 31 (11) (2019) 4857–4868.
[63] J. Pirgazi, M.M. Pourhashem Kallehbasti, A. Ghanbari Sorkhi, An end-to-end deep
learning approach for plate recognition in intelligent transportation systems, Wirel.
Commun. Mob. Comput. 2022 (2022) 1–13.
[64] D. Jin, et al., A survey of community detection approaches: From statistical
modeling to deep learning, IEEE Trans. Knowl. Data Eng. (2021).
[65] C. Wang, et al., Deep neighbor-aware embedding for node clustering in attributed
Sajad Ahmadian received his B.S. degree in computer engi­
graphs, Pattern Recognit. 122 (2022) 108230.
neering from Razi University, Kermanshah, Iran, in 2011, and
[66] J. Xie, R. Girshick, A. Farhadi, Unsupervised deep embedding for clustering
the M.Sc. degree in computer engineering, artificial intelli­
analysis, Int. Conf. Mach. Learn. (2016).
gence from University of Kurdistan, Sanandaj, Iran, in 2014.
[67] Z. El Yebdri, et al., Context-aware recommender system using trust network,
Moreover, he has been received his Ph.D. degree in computer
Computing 103 (9) (2021) 1919–1937.
engineering, artificial intelligence from University of Zanjan,
[68] S. Yadav, et al., Trust aware recommender system using swarm intelligence,
Zanjan, Iran, in 2018. He conducted a part of his Ph.D. research
J. Comput. Sci. 28 (2018) 180–192.
work in the laboratory of complex networks, RMIT University,
[69] Kaitlin Woolley, Ayelet Fishbach, A recipe for friendship: Similar food
Melbourne, Australia, form February 2018 to July 2018. He is
consumption promotes trust and cooperation, J. Consum. Psychol. 27 (1) (2017)
currently an Assistant Professor at the Faculty of Information
1–10, https://doi.org/10.1016/j.jcps.2016.06.003.
Technology, Kermanshah University of Technology, Kerman­
[70] B, De Finetti. Theory of Probability: A Critical Introductory Treatment, John Wiley
shah, Iran. His research interests include recommender sys­
& Sons, 2017.
tems, social networks analysis, data mining, and machine
[71] X. Qi, et al., Laplacian centrality: a new centrality measure for weighted networks,
learning.
Inf. Sci. 194 (2012) 240–253.
[72] WHO, Population Nutrient Intake Goals for Preventing Diet-related Chronic
Diseases. Diet, Nutrition and the Prevention of Chronic Diseases, 916, 2003, pp.
46–83. Vahid Farrahi received the Ph.D. degree in medical physics
and technology from the Faculty of Medicine, University of
Oulu, Finland, in 2021. He received funding from the European
Mehrdad Rostami holds a B.Sc. degree (Hons.) in computer Union’s Horizon 2020 Research and Innovation Program under
engineering from Razi University, Kermanshah, Iran, in 2012, the Marie Skłodowska-Curie Grant to perform his doctoral
and an M.Sc. degree (Hons.) in artificial intelligence from the research. His doctoral thesis received multiple recognition and
University of Kurdistan, Sanandaj, Iran, in 2014. Currently, he was selected as the Best Doctoral Thesis 2021 at the Faculty of
is a Ph.D. Researcher at the University of Oulu. His research Medicine, University of Oulu. From 2021 to 2023, he served a
focuses on various areas, including pattern recognition, ma­ Postdoctoral Fellow within the same research unit at the Uni­
chine learning, recommender systems, big data mining, and versity of Oulu. He is currently an Assistant Professor at the
digital health. Institute for Sport and Sport Science, TU Dortmund University,
Dortmund, Germany, where he conduct multidisciplinary
problem-oriented research, with a primary focus on machine
learning and data-driven solutions.

Mourad Oussalah received the Ph.D. degree in robotics and

artificial intelligence from the University of Paris XII, in 1998.
He held the research and academic position at KU Leuven, City
Kamal Berahmand received an M.S. degree in information
University of London, and Birmingham University, U.K. He is
technology engineering from Azarbaijan Shahid Madani Uni­
currently working as a Research Professor with the Faculty of
versity, Tabriz, Iran, in 2017. He is currently a Ph.D. student
ITEE, University of Oulu, Finland. He worked extensively on
with the Faculty of Computer Engineering and Information
information processing, data mining, and text mining, where
Technology, Queensland University of Technology, Brisbane,
he published more than 200 technical papers and led several
Australia. His current research interests include machine
projects. He is a fellow of the Royal Statistical Society.
learning, network science, graph neural network, and knowl­
edge graph.

15