A Novel Explainable and Health-aware Food Recommender System
Merhrdad Rostami
1 a
, Vahid Farahi
1,2 b
, Kamal Berahmand
3 c
, Saman Forouzandeh
4 d
,
Sajad Ahmadian
5 e
and Mourad Oussalah
1,2 f
1
Center for Machine Vision and Signal Analysis (CMVS), University of Oulu, Finland
2
Research Unit of Unit of Medical Imaging, Physics and Technology, University of Oulu, Finland
3
School of Computer Sciences, Queensland University of Technology, Australia
4
School of Mathematics and Statistics, Faculty of Science, University of New South Wales (UNSW), Australia
5
Faculty of Information Technology, Kermanshah University of Technology, Iran, Islamic Republic of
Keywords:
Food Recommender System, Healthy Recommendation, Explainability, Time-aware Recommendation.
Abstract:
Food recommendation systems are increasingly being used by online food services to make recommendations.
Health factors are often ignored in most of these systems, despite the fact that unhealthy diets are connected
to a wide range of non-communicable diseases. Furthermore, if users do not receive compelling explanations
about the recommended healthy foods, they may become hesitant to try them. In this paper, a novel explainable
and health-aware food recommender system is developed to address these challenges. For this purpose, user’s
preferences and food health factors are taken into account simultaneously and then a rule-based mechanism is
employed for final healthy and explainable recommendations. Five performance metrics were used to compare
our system with different new recommender systems. Using a dataset crawled from ”Allrecipes.com”, the
proposed model is shown to perform best.
1 INTRODUCTION
The rapid development of online food services
yielded several prototypes of food recommendation
systems that can assist users in finding appropriate
foods according to their preferences (Shaikh et al.,
2022; Ghosh et al., 2021). Despite the fact that previ-
ous food recommendations achieved acceptable effi-
ciency in terms of learning user’s preferences by map-
ping historical interactions with foods, these mod-
els still suffer from two significant limitations. The
first one is associated with the multi-objective na-
ture of healthy food recommendation. Recommend-
ing foods based on a target user’s diet preferences and
healthy food recommendations are two of the main
objectives of a healthy food recommender system. In
many cases, these two goals are in conflict with each
a
https://orcid.org/0000-0001-5710-217X
b
https://orcid.org/0000-0001-8355-8488
c
https://orcid.org/0000-0003-4459-0703
d
https://orcid.org/0000-0002-5952-156X
e
https://orcid.org/0000-0001-8355-8488
f
https://orcid.org/0000-0002-4422-8723
other, and maximizing one of these two objectives
may hurt the second objective. The second major
limitation/challenge of food recommender systems is
the lack of explainability and transparency. People
may be reluctant to try healthy foods as well as dis-
couraged to follow the recommendations if they do
not receive compelling explanations for the healthy
food recommendations they receive. For this rea-
son, a real and efficient healthy food recommendation
is one that explains to each user why the food was
recommended for him/her or why another unhealthy
food that may be more appropriate for his/her pref-
erences was not recommended. We have addressed
the above-mentioned limitations in this paper by de-
veloping a novel food recommender system based
on health-aware multi-objective function and explain-
able/controllable recommendation. In comparison to
previously developed food recommendation systems,
the developed system makes significant contributions,
which are outlined below.
• Multi-objective: By integrating health and nutri-
tion factors into the food recommendation frame-
work, this study guides users toward a healthy eat-
208
Rostami, M., Farahi, V., Berahmand, K., Forouzandeh, S., Ahmadian, S. and Oussalah, M.
A Novel Explainable and Health-aware Food Recommender System.
DOI: 10.5220/0011561700003335
In Proceedings of the 14th International Joint Conference on Knowledge Discovery, Knowledge Engineering and Knowledge Management (IC3K 2022) - Volume 3: KMIS, pages 208-215
ISBN: 978-989-758-614-9; ISSN: 2184-3228
Copyright
c
 2022 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

ing style. The study incorporates a multi-objective
measure to consider user preferences and health
simultaneously.
• Explainable Food: An effective explainable food
recommender system is introduced in this study.
To the best of our knowledge, this is the first ex-
plainable food recommender systems.
• Controllable: A novel controllable food recom-
mender system is developed that will enable the
target user to take part in the recommendation pro-
cess and strike a balance between the target users’
own preferences and the health of the food.
• Time and Ingredients-aware Similarity: A
novel ingredient and time-aware similarity mea-
sure is introduced to consider food contents and
the temporal information of ratings in user simi-
larity calculation.
• Dynamic Neighbor Selection: In this paper,
Contrary to the previous recommender system
(Ahmadian et al., 2022b; Ahmadian et al., 2022a),
by incorporating the ”friend of a friend” idea, a
novel dynamic transitive-based nearest neighbor
selection is developed.
The rest of this paper is organized as follows. We
review the related works in Section 2. The developed
recommender system is represented in Section 3. Ex-
perimental results are provided in Section 4. Finally,
Section 5 provide the conclusion.
2 LITERATURE REVIEW
This section reviews the previous food recommenda-
tion models. Then these proposed models, their short-
comings, and drawbacks are investigated.
In (Gao et al., 2022) a food recommendation using
a Graph Convolutional Network (FGCN), which uses
several embedding propagation layers to model high-
order connections and improve representation learn-
ing is developed. In (Gao et al., 2019), a novel hier-
archical attention-based food recommender system is
developed that takes into account user history about
nutrition, ingredients of a food. In (Asani et al.,
2021), a method for extracting customer food pref-
erences from online restaurant reviews is developed.
Their method uses NLP techniques to process the text
of user comments and extract appropriate food-related
terminology. The authors of (Shabanabegum et al.,
2020) developed a new model to suggest different
foods to the users based on the available food items
in the refrigerator.
Additionally to learning user preferences, a num-
ber of previous food recommendation systems con-
sider health issues and nutritional requirements. El-
sweiler et al. (Elsweiler et al., 2017) presented a
novel system for healthy food recommendation using
identifying the ingredients of the foods that received a
high score from the users. Bianchini et al. (Bianchini
et al., 2017) developed a novel method to personal-
ize the menu of food orders. In (Chen et al., 2020),
the authors presented a novel framework called Nu-
tRec that provides a guide to healthy eating. Further-
more, in (Forouzandeh and Aghdam, 2019), consid-
ering the behavior and lifestyle of users and their nu-
trition, a health recommender system is presented and
it provides recommendations for maintaining health.
Meng et al. (Meng et al., 2020) proposed a privileged-
channel infused network (PiNet) framework in a het-
erogeneous graph. This recommender system helps
users maintain a healthy diet by creating nutritional
memories. In another study (Wang et al., 2021),
the authors presented Market2Dish, a personalized
health-aware food recommendation scheme that can
identify each food’s ingredients and determine the im-
pact of nutrition on users’ health.
3 DEVELOPED SYSTEM
This section details the developed method as a novel
Explainable and Health-aware Food Recommender
System (in short, EHFRS), which is organized in
five main steps: (1) Time- and Ingredient-aware User
Similarity Calculation, (2) Neighbor Selection, (3)
Initial User-Preference Prediction, (4) Food Health
Factor Calculation, (5) Multi-Objective and Control-
lable Food Rating and (6) Explainable Food Rec-
ommendation. In the first step, a new time-aware
user similarity measure is introduced based on the
ingredients of rated foods by each user. A new dy-
namic neighbor selection mechanism is developed in
the second step. In the third step, the initial user-
preference for each food is predicted, while in the
fourth step, the health factor is evaluated for each food
item considering the ideal range of dietary factors
suggested by WHO. In the fifth step, based on user-
preference and estimated health factor, a new multi-
objective and controllable food rating function is in-
troduced. Finally, in the sixth step, the top healthy
preferred foods along with the explanations are rec-
ommended to the target user. Figure 1 illustrates the
general schema of the developed EHFRS system. The
details of the proposed system are described in the fol-
lowing subsections.
A Novel Explainable and Health-aware Food Recommender System
209

Figure 1: Overall schema of the developed EHFRS.
3.1 User Similarity Calculation
Most previous recommender systems rely mainly on
direct user ratings to calculate user similarity. When
such a system contains many items, while only a
small number of these items have been rated by
each user, the system accuracy becomes greatly re-
duced. This is especially true for food social networks
where different types of recipes are shared for each
food item, which renders the user-food matrix is very
sparse. To illustrate the inefficiency of the previous
similarity calculation measure, we provide an exam-
ple in Figure 2. In this example, food f
1
and food f
2
were rated 5 and 4, respectively, by user u
1
. On the
other hand, user u
2
has rated 4 and 5 ranks to food f
3
and food f
4
, correspondingly. It is apparent, at first
glance, that foods f
1
and f
3
as well as f
1
and f
3
are
very similar and differ only in one ingredient. Al-
though users u
1
and u
2
have similar diet tastes and
Figure 2: Simple example of user-ingredient interaction.
preferences, by traditional similarity calculation mea-
sures, their similarity will be calculated as 0, since
these users have not rated a common food.
In addition, each user’s food preferences may
change over time, so the voting time recorded by each
user must also be taken into account. Therefore, the
user similarity criterion will handle the importance of
time for different ratings where the old ratings will
have a lower importance than the new ratings. To
defeat this issue, a novel time- and ingredient-aware
measure is developed to calculate user similarities.
For this purpose, the time weight of users’ u
i
and u
j
ratings to food f
i
is defined by considering the time
stamp of those ratings. This time weight is deter-
mined as follows:
TW (u,v, f
i
) =
p
e
−λ(T P−t(u, f
i
))
× e
−λ(T P−t(v, f
i
))
. (1)
where, t (u, f
i
) denotes the time period of the regis-
tered rate of user u to food f
i
, T P indicates the max-
imum Time Period, and λ denotes a user control pa-
rameter that adjusts the impact of the time factor. A
high (resp. low) value of λ indicates a greater (resp.
smaller) impact of time factor in calculating similar-
ity values. According to the user’s time intervals, the
ratings are divided into different time periods. For
the context of our experiment detailed in the experi-
ment section, since the collected ratings span over a
period of 18 years, we divided the user ratings into
monthly intervals for our experiments. In the case of
more dense user ratings, weekly or even daily time
frames are potential alternatives. Moreover, consider-
ing the list of ingredients of each food, the similarity
between food f
i
and food f
j
is calculated as below:
SimF ( f
i
, f
j
) =
∑
K
k=1

ing
ik
× ing
jk

Max (
∑
K
k=1
ing
ik
,
∑
K
k=1
ing
jk
)
. (2)
where ing
ik
is an indicator variable indicating to the
availability of ingredient k in food f
i
; namely, if food
KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems
210

f
i
contains ingredient k, ing
ik
will be set to 1, oth-
erwise, ing
ik
= 0. After calculating the time weight
and food similarities, the user similarity sim (u, v) be-
tween user u and user v is calculated as below:
Sim(u,v) =
a
bc
(3)
where







a =
∑
f
i
∈F
∑
f
j
∈F
((r
i
(u) − ¯r (u)) × (r
j
(v) − ¯r (v)) × SimF( f
i
, f
j
) × T W (u,v, f
i
)),
b =
q
∑
f
i
∈F
(r
i
(u) − ¯r (u))
2
× SimF( f
i
, f
j
) × T W (u,v, f
i
),
c =
q
∑
f
j
∈F
(r
j
(u) − ¯r (u))
2
× SimF( f
i
, f
j
) × T W (u,v, f
j
).
where r
i
(u) is the rating given to food f
i
by user u,
and ¯r (u) is the average rating given by user u, and
F is the set of initial foods in the system. Moreover,
TW (u, v, f
i
) and SimF( f
i
, f
j
) denote the Weight Time
of the ratings of users’ rates u and v for food f
i
and
Food similarity between Food f
i
and f
j
, correspond-
ingly, calculated using Eq. 1 and Eq. 2, respectively.
3.2 Neighbor Selection
The collaborative filtering-based recommender sys-
tems’ primary objective is to identify nearest neigh-
bors through the rating prediction process and to cal-
culate the final recommendations. Due to data spar-
sity of the user rating data, finding neighbors in a nor-
mal way may produce inefficient information. There-
fore, in this study, a transitive-based nearest neighbor
selection is developed. By combining the ”friend of a
friend” idea, we developed a new method for nearest
neighbor selection, which includes all k-order near-
est neighbors. In this proposed method, (k-1)-order
neighbors will be used to calculate k-order neighbors.
Our method incrementally expands neighbors, which
would prepares reliable neighborhood knowledge for
the final recommendation.
Let N
1
u, f
i
be a set of users that have rated food f
i
.
Then, the k-order nearest neighbors of user u for the
food f
i
is calculated as follows:
N
k
u, f
i
=
n
N
k−1
u, f
i
∪ Neighbors(u)
o
. (4)
where Neighbors(u) is defined as below:
Neighbors(u) =
{
w ∈ U : Sim(u,w) > θ
}
. (5)
where, Sim(u,w) indicates the user similarity between
user u and w computed using Eq 3, U is the set of
all users in the system, and θ is a threshold parame-
ter for neighbors’ selection. Finally, the k-order near-
est neighbors of user u (regardless of the food item)
is calculated by taking into account all food items as
follows:
SumN(u)
k
=
[
f
i
∈F
N
k
u, f
i
. (6)
3.3 Initial User-preference Prediction
In this step, the users’ neighbors selected are utilized
to predict the food preferences of users. Let N
k
u, f
i
be a
set of k-order nearest neighbors of user u for food f
i
.
Then, the predicted rating P
i
(u) of an unknown food
f
i
for user u is given as follows:
P
i
(u) = ¯r
u
+
∑
v∈N
k
u, f
i
Sim(u,v) × (r
v,i
− ¯r
v
)
∑
v∈N
k
u, f
i
Sim(u,v)
. (7)
where ¯r
u
corresponds to the average ratings assigned
by user u, r
v,i
is the rating of food f
i
assigned by user
v, and Sim(u, v) refers to the similarity score between
users u and v calculated using Eq 3.
3.4 Food Health Factor Calculation
Food recommendation systems are more critical than
recommendation systems associated to movie, mu-
sic or book because of the user’s health impact of
the underlined recommendation. Accordingly, assess-
ing how healthy the recommendation is is critical in
the creation of efficient food recommendation where
healthy diets are significant in preventing and treating
chronic diseases. Furthermore, several expert groups
have recognized that diet is a critical contributor to
noncommunicable disease etiology, highlighting the
need to change eating habits (Organization et al.,
2014). WHO has encouraged the use of nutrient pro-
filing tools to encourage healthy diet and reduce the
burden of non-communicable diseases (Lawlor and
Pearce, 2013). Therefore, we have used the amount of
macro-nutrients as a performance metric to assess the
health factor of a given food. This factor evaluates the
amount of seven types of nutrition categories includ-
ing proteins, carbohydrates, sugars, sodium, fat, satu-
rated fats, and fibers as per WHO suggestion. WHO
provided an appropriate range for each of these nu-
trition categories, so that the food is deemed healthy
if the amount of each category fails withing the pro-
vided range (Table 1) (Organization et al., 2007). Us-
ing the WHO ideal range, the Health Factor of food f
i
can be defined as follows:
HF ( f
i
) = Pro( f
i
) +Carb( f
i
) + Sug ( f
i
)
+Sod ( f
i
) + Fat ( f
i
) + Sat ( f
i
) + Fib ( f
i
).
(8)
where Pro ( f
i
), Carb( f
i
), Sug ( f
i
), Sod ( f
i
), Fat ( f
i
),
Sat ( f
i
) and Fib ( f
i
) indicate proteins, carbohydrates,
sugars, sodium, fat, saturated fats, and fibers of food
f
i
are in ideal range or not, respectively. For ex-
ample, Pro ( f
i
) = 1 if the amount of the protein
in food f
i
is in the ideal range shown in Table 1,
otherwise; Pro( f
i
) = 0. Therefore, it can be con-
cluded that the value of HF ( f
i
) is in the range of
[0(unhealthy), 7 (healthy)].
A Novel Explainable and Health-aware Food Recommender System
211

Table 1: Ideal ranges of nutrition(Organization et al., 2007).
Dietary factor Ideal Range
Proteins 10-15%
Carbohydrates 55-75 %
Sugars <10 %
Sodium <5 g
Fats 15-30%
Saturated fats <10 %
Fibers >10 g
3.5 Multi-objective Food Rating
An efficient and healthy food recommender system
must consider the food recommendation as a multi-
objective problem by balancing quality metrics such
as the preferences of users and the health factor.
After predicting the user preferences and healthy
factor of the foods, the final rating of food f
i
for user
u can be predicted as follows:
FP
i
(u) = (1 − γ).P
i
(u) + γ.HF( f
i
) . (9)
where P
i
(u) is the predicted user preference of user u
for food f
i
calculated using Eq. (7), HF( f
i
) is Healthy
Factor content of food f
i
calculated using Eq. (8).
While the parameter γ balances between user prefer-
ence objective and the health factor objective. This
parameter ranges from 0 to 1. With a higher γ param-
eter, food health objective becomes more significant
in the final recommendation. The ability to give users
control over a multi-objective recommender system
allows the users to have an direct effect on the fi-
nal recommendations, i.e., they can filter or re-sort
the recommendations based on their preferences and
healthy factor of foods. Typically, an interactive visu-
alization framework is necessary to support user inter-
action during the recommendation process. This con-
trollable parameter will allow users to participate in
the final food recommender system by deciding which
of these two goals (i.e., preference and healthy) is
most important to them. In other words, by providing
a user parameter, i.e., adjusting the health factor pa-
rameter, our controllable food recommender system
lets end-users become part of the food recommenda-
tion process.
3.6 Explainable Food Recommendation
Users usually have little understanding of how the
system comes up with healthy recommendations, so
the reasons for receiving them remain opaque. The
aim of this paper is to propose an explainable healthy
food recommendation system that can explain why
some unhealthy foods are not recommended and some
healthy foods are. Explainable recommender sys-
tems may either be incorporated as part of their in-
herent design (intrinsic explainability) or provided as
a post-doc explainable addition to the recommender
systems(Rostami and Oussalah, 2022b; Rostami and
Oussalah, 2022a). A new explainable rule mining
model is developed for the final food recommenda-
tions in this step. During this step, the aim is to deter-
mine the rules in the form of f
i
→ f
j
, meaning that if
a user has previously tasted (or liked) food f
i
, he/she
will also be interested in food f
j
. In order to iden-
tify these rules, selected neighbors for each user and
multi-objective food ratings are considered. For this
purpose, let F
u
be the set of foods rated by the target
user u and let F
N
u
be the set of foods rated by nearest
neighbors of the user u. Next, the set of foods rated
by all the users in SumN(u)
k
except the target user u,
which has not rated them yet, is defined as below:
F
0
u
= F
SumN(u)
k
− (F
u
∩ F
SumN(u)
k
). (10)
Then, for each food f
i
∈ F
u
and each food f
j
∈ F
0
u
the confidence value of rule f
i
→ f
j
can be calculated
using the following equation:
con f ( f
i
→ f
j
) =
n( f
i
, f
j
)
n( f
i
)
. (11)
where n( f
i
) is the number of users in SumN(u)
k
who
have rated food f
i
, and n( f
i
, f
j
) is the number of users
in SumN(u)
k
who have rated both food f
i
and f
j
.
Then, the preference rank of food f
j
∈ F
0
u
for the
target user u is calculated as follows:
Pre f Rank
j
(u) = arg max
f
i
∈F
u
(con f ( f
i
→ f
j
) × P
j
(u)).
(12)
where P
j
(u) indicates the predicted rating of food f
j
for user u calculated using Eq. (7).
Accordingly, the healthy-preference rank of food
f
j
∈ F
0
u
for target user u is defined as below:
HealthPre f Rank
j
(u) = arg max
f
i
∈F
u
(con f (i → j)×FP
j
(u)).
(13)
where FP
j
(u) is the final healthy and preference rat-
ing of food f
j
for user u calculated using Eq. (9).
Then, according to the obtained preference rank
and healthy-preference rank of foods, we can iden-
tify the preference and healthy-preference foods. Let
the Preference Food set of user u be Pre f F(u),
where unseen foods for user u are the Top-L high-
est Pre f Rank
j
(u) based on Eq 12. Moreover, the
healthy-preference food set of user u, HealthF(u) are
the Top-L highest ranks HealthPre f Rank
j
(u) accord-
ing to Eq 13. Therefore, the set of Unhealthy and
KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems
212

Preference foods for user u can be calculated as fol-
lows:
UP(u) =
{
f ∈ F| f ∈ Pre f F(u) ∧ f /∈ HealthF(u)
}
.
(14)
The set of Unhealthy and Preference foods for user u
UP(u) will be introduced to the target user with this
explanation:
While these foods may be your favorites, they are
not recommended due to their unhealthy content.
Moreover, for each food in the HealthF(u) food
set, considering the rules ( f
i
→ f
j
) according to Eq
13 that leads to this recommendation where the fol-
lowing explanation is displayed to the user:
This food is also of interest to users who have
liked food f
i
.
4 EXPERIMENTAL RESULTS
We conducted extensive experiments to assess the
effectiveness of the developed EHFRS. As part
of the evaluation of our model, we crawled the
www.Allrecipes.com food social network, and ex-
tracted 52,821 foods for the period 2000-2018. Users’
ratings, food nutrition, and timestamps are crawled
for each food (Gao et al., 2019). The rating of a
variety of foods is used to generate an implicit feed-
back, denoting whether the users interacted with food
items. Totally, 68,768 users, 45,630 food items and
1,093,845 ratings were obtained.
In order to assess the effectiveness of the de-
veloped food recommender system, five well-known
metrics have been utilized, including Precision, Re-
call, F1, AUC, and NDCG. There is no straightfor-
ward way to evaluate the precision and recall of rec-
ommender systems because each item needs to be
rated by the user to determine if it is relevant. To this
end, in our experiments, we employ Precision@N,
Recall@N, and F1@N (N=size of the recommenda-
tion list).
4.1 Experimental Setup
The developed EHFRS model has four input param-
eters, which their values should be initialized before
performing the experiments. The first parameter is λ,
which specifies the importance of time factor in cal-
culating time-aware similarity values (Eq. (1)). The
parameter of k is the second parameter that denotes
the order of order nearest neighbors in neighbor selec-
tion step (Eq. (4)). Moreover, θ is the third parameter
and used in the neighbor definition formula (Eq. (5)).
Finally, γ is the fourth parameter to balance between
Figure 3: Performance analysis of Time factor.
user preference objective and the health factor objec-
tive (Eq. (9)). Accordingly, the values of λ, k and θ
parameters are set to 4, 0.6 and 3, respectively. More-
over, the sensitivity analysis of γ parameter is shown
in further experiments.
4.2 Performance Comparison
We conduct extensive experiments to verify the ef-
fectiveness of the developed EHFRS model. We
then report the results and discuss them using vari-
ous aforementioned evaluation metrics. In order to
make a comparison, four state-of-the-art food rec-
ommendation approaches including Hierarchical At-
tention Food Recommendation (HAFR) (Gao et al.,
2019), Collaborative Filtering Recipe Recommenda-
tions (CFRR) (Chavan et al., 2021), Heterogeneous
Recipe Graph Recommendation model(HGAT) (Tian
et al., 2021) and Food recommendation with Graph
Convolutional Network (FGCN) (Gao et al., 2022) are
chosen as baselines.
In the first part of our experiments, we will in-
vestigate the effect of taking timestamps into account
when predicting user rates by our developed food
recommender system. Figure 3 evaluates the effi-
ciency of the proposed food recommendation with
and without taking the timestamps into account when
recommendation process. As shown in this figure,
the developed time-aware food recommender sys-
tem predicts ratings substantially better than its time-
unaware counterpart. For example, Precision@10,
Recall@10, F1@10 and NDCG@10 improved by
21,13%, 20.72%, 18.13% and 11.38, respectively.
In the next experiment, the different food rec-
ommender systems are compared in terms of Preci-
sion@10, Recall@10, F1@10, AUC and NDCG@10.
Table 2, shows the performance of different food rec-
ommender systems. The developed food recommen-
dation model (i.e., EHFRS) outperformed all other
state-of-the-art models based on all of the evalua-
tion metrics. Moreover, the study of these results
A Novel Explainable and Health-aware Food Recommender System
213

Table 2: Performance of compared food recommendations.
Method Precision Recall F1 AUC NDCG
HAFR 0.0692 0.0671 0.0687 0.6439 0.0451
CFRR 0.0671 0.0647 0.0637 0.6421 0.0431
HGAT 0.0672 0.0649 0.0638 0.6431 0.0436
FGCN 0.0710 0.0681 0.0695 0.6639 0.0462
EHFRS 0.0739 0.0703 0.0717 0.6884 0.0512
reveals that the proposed method is 4.08%, 3.23%,
3.16%, 3.69% and 10.82% more efficient than the
second-best food recommender system (i.e., FGCN)
in terms of Precision@10, Recall@10, F1@10, AUC
and NDCG@10 metrics, respectively. It should be
noted that in this experiment the γ parameter of the
developed EHFRS that controls the food healthy fac-
tor’s of recommendations is set to 0.2.
The next experiments investigate the effects of
changing the size of the recommendation list on the
different metrics. The experiments examined the per-
formance of the different food recommender systems
when the recommendation list sizes were 10, 15, and
20. Figures 4 - 6 illustrate the effects of the size of the
recommendation list on Precision, Recall, and NDCG
metrics. It was found that increasing the size of
the recommendation list increases Recall and NDCG
Figure 4: Precision of Top-N recommended foods.
Figure 5: Recall of Top-N recommended foods.
Figure 6: NDCG of Top-N recommended foods.
metrics and reduces Precision. Furthermore, the re-
ported results in these experiments demonstrated that
the developed food recommender system consistently
outperformed other compared systems.
The health controllable parameter of γ, which de-
termines how much healthy factor of foods is con-
sidered in the final recommendations, is one of the
most important parameters of the developed food rec-
ommender system. A user can adjust this parameter
based on how important the health of the food is to
him/her. It is clear that the more this parameter is set
to a high value, the more important the food health
will be in the food recommendation process, and as
a result, these recommendations will be further away
from the user’s preferences. In fact, setting this pa-
rameter to a high value can reduce the classic metrics
of the recommender system, including Precision, Re-
call, F1, and NDCG. In Figure 7, the effect of γ pa-
rameter, when it increases from 0 to 0.8, on the differ-
ent recommended systems metrics is investigated. As
the results show, changing the Delta parameter from
zero to 0.8 reduced the Precision@10, Recall@10,
F1@10, and NDCG@10 metrics by about 52.02%,
37.73%, 45.58% and 40.09%, respectively.
Figure 7: Performance evaluation of the health factor.
KMIS 2022 - 14th International Conference on Knowledge Management and Information Systems
214

5 CONCLUSION
The use of food recommendation systems has sig-
nificantly increased in online food services in recent
years, aiming at providing users with personalized
food recommendations. In this paper, a novel ex-
plainable and health-aware food recommender system
is developed and its performance is investigated on
the real-world food social network. In terms of Pre-
cision, Recall, F1, AUC, and NDCG, the developed
EHFRS performed significantly better than state-of-
the-art food recommendation models.
In addition, traditional food recommendation
models generally focus on single-user scenarios, how-
ever, most real-life interactions take place in groups,
In future, we plan to developed food group recom-
mendation models. In addition, in future, we aim to
enhance the performance of the food recommendation
by incorporating additional user information.
ACKNOWLEDGEMENTS
The project is supported by the Academy of Fin-
land (project number 326291) and the University of
Oulu Academy of Finland Profi5 on Digihealth. This
work also was supported in part by the Ministry of
Education and Culture, Finland (OKM/20/626/2022).
Moreover, SA was supported by the Kermanshah
University of Technology, Iran, under grant number
S/P/F/5.
REFERENCES
Ahmadian, M., Ahmadi, M., and Ahmadian, S. (2022a). A
reliable deep representation learning to improve trust-
aware recommendation systems. Expert Systems with
Applications, 197:116697.
Ahmadian, S., Joorabloo, N., Jalili, M., and Ahmadian, M.
(2022b). Alleviating data sparsity problem in time-
aware recommender systems using a reliable rating
profile enrichment approach. Expert Systems with Ap-
plications, 187:115849.
Asani, E., Vahdat-Nejad, H., and Sadri, J. (2021). Restau-
rant recommender system based on sentiment analy-
sis. Machine Learning with Applications, 6:100114.
Bianchini, D., De Antonellis, V., De Franceschi, N., and
Melchiori, M. (2017). Prefer: A prescription-based
food recommender system. Computer Standards &
Interfaces, 54:64–75.
Chavan, P., Thoms, B., and Isaacs, J. (2021). A recom-
mender system for healthy food choices: Building a
hybrid model for recipe recommendations using big
data sets. In Proceedings of the 54th Hawaii Interna-
tional Conference on System Sciences, page 3774.
Chen, M., Jia, X., Gorbonos, E., Hoang, C. T., Yu, X., and
Liu, Y. (2020). Eating healthier: Exploring nutrition
information for healthier recipe recommendation. In-
formation Processing & Management, 57(6):102051.
Elsweiler, D., Trattner, C., and Harvey, M. (2017). Exploit-
ing food choice biases for healthier recipe recommen-
dation. In Proceedings of the 40th international acm
sigir conference on research and development in in-
formation retrieval, pages 575–584.
Forouzandeh, S. and Aghdam, A. R. (2019). Health rec-
ommender system in social networks: A case of face-
book. Webology, 16(1).
Gao, X., Feng, F., He, X., Huang, H., Guan, X., Feng, C.,
Ming, Z., and Chua, T.-S. (2019). Hierarchical at-
tention network for visually-aware food recommenda-
tion. IEEE Transactions on Multimedia, 22(6):1647–
1659.
Gao, X., Feng, F., Huang, H., Mao, X.-L., Lan, T., and Chi,
Z. (2022). Food recommendation with graph convolu-
tional network. Information Sciences, 584:170–183.
Ghosh, P., Bhattacharjee, D., and Nasipuri, M. (2021). Dy-
namic diet planner: A personal diet recommender sys-
tem based on daily activity and physical condition.
IRBM, 42(6):442–456.
Lawlor, D. A. and Pearce, N. (2013). The vienna declaration
on nutrition and non-communicable diseases.
Meng, L., Feng, F., He, X., Gao, X., and Chua, T.-S. (2020).
Heterogeneous fusion of semantic and collaborative
information for visually-aware food recommendation.
In Proceedings of the 28th ACM International Confer-
ence on Multimedia, pages 3460–3468.
Organization, W. H. et al. (2007). Population nutrient in-
take goals for preventing diet-related chronic diseases.
http://www. who. int/nutrition/topics/.
Organization, W. H. et al. (2014). Global status re-
port on noncommunicable diseases 2014. Number
WHO/NMH/NVI/15.1. World Health Organization.
Rostami, M. and Oussalah, M. (2022a). Cancer predic-
tion using graph-based gene selection and explainable
classifier. Finnish Journal of eHealth and eWelfare,
14(1):61–78.
Rostami, M. and Oussalah, M. (2022b). A novel explain-
able covid-19 diagnosis method by integration of fea-
ture selection with random forest. Informatics in
Medicine Unlocked, 30:100941.
Shabanabegum, S., Anusha, P., Seethalakshmi, E., Shun-
mugam, M., Vadivukkarasi, K., and Vijayakumar, P.
(2020). Iot enabled food recommender with nir sys-
tem. Materials Today: Proceedings.
Shaikh, S. G., Suresh Kumar, B., and Narang, G. (2022).
Recommender system for health care analysis using
machine learning technique: a review. Theoretical Is-
sues in Ergonomics Science, pages 1–30.
Tian, Y., Zhang, C., Metoyer, R., and Chawla, N. V. (2021).
Recipe recommendation with hierarchical graph at-
tention network. Frontiers in Big Data, 4:1–13.
Wang, W., Duan, L.-Y., Jiang, H., Jing, P., Song, X., and
Nie, L. (2021). Market2dish: Health-aware food
recommendation. ACM Transactions on Multime-
dia Computing, Communications, and Applications
(TOMM), 17(1):1–19.
A Novel Explainable and Health-aware Food Recommender System
215