Food Recommendation in a Worksite Canteen

Vincenza Carchiolo

1 a

, Marco Grassia

2 b

, Alessandro Longheu

2 c

Michele Malgeri

2 d

and Giuseppe Mangioni

2 e

Dip. Matematica e Informatica, Universit

a di Catania, Italy

Dip. Ingegneria Elettrica Elettronica Informatica, University of Catania, Italy

Keywords:

Data Analysis, Recommendation System, Machine Learning.

Abstract:

Recommendation systems tackle with information overload to assist people in ﬁnding their best choice accord-

ing to their preferences and past behaviour. This occurred in many contexts, including the food sector where

culinary inspiration, sales increase or healthy advice motivate the adoption of such a system. In this paper we

propose a canteen food recommendation system for workers operating at an innovation hub including more

than 20 companies. The system leverages a 30 months data set of past choices, and adopts a content based

and a collaborative ﬁltering approach for canteen users, suggesting them with dishes chosen by other similar

users. First results for frequent as well as occasional canteen visitors are encouraging to validate the proposed

approach.

1 INTRODUCTION

Information overload in decision-making processes

exploits recommendation systems (Bobadilla et al.,

2013) (Mohamed et al., 2019) both as a tool to help

users ﬁnd products based on their preferences and

past choices, and to assist companies in making tar-

geted sales based on customers’ interests; one of the

areas where such systems are widely used is the rec-

ommendation of meals (Min et al., 2019).

The food sector is important for various reasons

(Nations, 2015)(Torreggiani et al., 2018), as improv-

ing culinary inspiration, increasing sales, and promot-

ing health improvement. Food recommendation has

been studied by many researchers (Jiang et al., 2019),

(Merler et al., 2016), (Iwendi et al., 2020); to create an

effective system it is crucial to understand how peo-

ple make food-related choices considering factors as

cultural, social, economic and even organic.

In this article we present an application that assists

a worker in booking his/her meal at a canteen operat-

ing in an innovation hub located in Italy (Carchiolo

et al., 2020), (Carchiolo et al., 2021). The canteen

menu manager is a native Italian speaker that inserts

https://orcid.org/0000-0002-1671-840X

https://orcid.org/0000-0001-5841-6058

https://orcid.org/0000-0002-9898-8808

https://orcid.org/0000-0002-9279-3129

https://orcid.org/0000-0001-6910-0112

in the menu local names for dishes, which are usually

non self-explanatory; this results in a certain discom-

fort for some workers when they have to choose. For

this reason, a key feature of the app is the engine of a

recommendation system that will suggest each worker

with the pot he/she will most likely to prefer. We col-

lected data concerning dishes consumed at a worksite

during about 2.5 years by more than 200 workers of

22 different companies. Starting from this dataset, we

set up a recommendation system aiming to suggest

users with dishes chosen by other similar users. We

adopted a content based approach and a collaborative

ﬁltering approach to manage both frequent and occa-

sional canteen visitors; ﬁrst results are encouraging to

validate our proposal.

In section 2 related works on food recommenda-

tion systems are considered, whereas in section 3 the

case study is introduced and results from dataset are

discussed, and ﬁnally in section 4 some conclusions

and proposals for future activities are presented.

2 FOOD RECOMMENDATION

Food Recommendation (FR) system are diversiﬁed

according to their data model, the number of sources

data extracted from and users interaction support;

they involve disparate factors (e.g. cultural prefer-

ences or medical prescriptions), in predicting what

Carchiolo, V., Grassia, M., Longheu, A., Malgeri, M. and Mangioni, G.

Food Recommendation in a Worksite Canteen.

DOI: 10.5220/0010502401170124

In Proceedings of the 6th International Conference on Complexity, Future Information Systems and Risk (COMPLEXIS 2021), pages 117-124

ISBN: 978-989-758-505-0

117

people prefer to eat (Elsweiler et al., 2017).

FR system can be used in several scenarios:

• Seasonality of ingredients, products and recipes

determined by analyzing historical trends, to help

in menu planning or to optimize purchase / inven-

tory decisions for restaurants;

• food Diets, using nutrition data to guide customer

preferences according to levels of (or sensitivity

to) salt, sugar, fat, etc.;

• increase in efﬁciency and cost savings in product

or recipe formulations by selecting cheaper equiv-

alents of current ingredients;

• customer Trends analysis, gathering which

recipes and products users ultimately act or buy,

and the time/season/event those recipes or prod-

ucts are associated to;

• recipe generator for food waste reduction. Start-

ing from available food in the fridge or pantry at

a given moment and using a recipes dataset, a re-

curring neural network can generate a complete

recipe that includes the instructions, the category

and the title (Zhang et al., 2021);

• meal recommendations based on user preferences

provided in the form of meal ratings, exploiting

the tastes of the individual user or the collabora-

tive approach, which is based on the relationships

between users or elements (meals).

2.1 Classiﬁcation of FR Systems

Based on the existing literature, we classify FR sys-

tems into three categories, i.e content-based, collabo-

rative ﬁltering and hybrid approaches.

Content-based approaches aim to tailor recom-

mendations to the user’s individual tastes. This is

achieved building a proﬁle for each user starting from

the attributes used to describe the characteristics of

the meals. (Freyne et al., 2011) introduce recommen-

dations by extracting from the recipes of the various

meals the individual ingredients and the positive eval-

uations of the users about those ingredients. This

means that if an ingredient is present in recipes that

a user ﬂagged as liked, other recipes containing the

same ingredient will also be suggested. Subsequent

work has pursued this approach taking into account

both positive and negative ingredient evaluations, rat-

ing recipes ingredients (Harvey and Elsweiler, 2017).

Other content-based approaches are speciﬁcally

suitable to FR systems. For example, since food

decisions are often visually guided, images associ-

ated with recipes are exploited (Zhang et al., 2020).

In (Yang et al., 2017) authors show that basic ap-

proaches can be outperformed exploiting (Elsweiler

et al., 2017) also show that low-level image functions

automatically extracted, such as brightness, color, and

sharpness can be useful for predicting a user’s food

preference.

Collaborative, ﬁlter-based methods for meal rec-

ommendation systems have been proposed, in partic-

ular in item-based collaborative ﬁltering recommen-

dations are based on how meals resemble each other,

whereas in user-based collaborative ﬁltering recom-

mendations are based on the preferences provided by

the user and their emerging similarity. In (Freyne and

Berkovsky, 2010) authors present an approach using

Pearson’s correlation on the classiﬁcation matrix, but

with a worse ranking than the content-based approach

described above. Harvey et al. (Harvey and Elsweiler,

2015) demonstrated that Singular Value Decomposi-

tion (SVD) outperformed both the content-based and

collaborative ﬁltering approaches. A Matrix Factor-

ization (MF) approach for food recommendation sys-

tems that merges classiﬁcation information and user-

supplied labels to achieve signiﬁcantly better predic-

tion accuracy than baselines based on decomposition

by content and standard matrices is presented (Ge

et al., 2015).

Among Hybrid ﬁltrations (Freyne and

Berkovsky, 2010) combined a collaborative user-

based ﬁltering method with a content-based method.

In their follow-up work involving user groups, the

same authors used an hybrid approach to combine

three different FR strategies into a single model based

on the ratio between the number of rated articles

respect to the total number. In (Harvey and Elsweiler,

2017) signiﬁcant results combining an SVD approach

with user and meal bias are achieved.

2.2 FR System Applications

Several commercial solutions for food recommenda-

tions are available with different strategies, as (1) sim-

ple user ratings or (2) user preferences extracted from

their past choices, or (3) similarity between meals

based on their ingredients/allergens, (4) similarity be-

tween users calculated by deﬁning a proﬁle for each

user based on past preferences, or ﬁnally (5) exploit-

ing nutritional needs in terms of calories. In the fol-

lowing, some applications are brieﬂy discussed.

Yum-me (Yang et al., 2017) is a FR system based

on personalized nutrition; it learns food preferences

without relying on the user’s dietary history. The

recommender learns users’ food preferences through

a simple visual quiz-based interface and then at-

tempts to generate meal recommendations that meet

the user’s health goals, food restrictions, and personal

appetite for food. It can be used by people who have

COMPLEXIS 2021 - 6th International Conference on Complexity, Future Information Systems and Risk

118

food restrictions, such as vegetarian, vegan, etc. and

it is based on two steps:

1. Users answer a simple survey to specify their

dietary restrictions and nutritional expectations.

This is used to ﬁlter foods and create an initial set

of candidates for recommendations.

2. Users then use an adaptive visual interface to ex-

press their food preferences through simple food

comparisons. The preferences learned are used to

further reﬁne the proposed recommendations.

Another example of food recommendation system

is Caviar (HBS-Digital-Initiative, 2020), a commer-

cial system that can be customized via an user se-

lected optimization function, for example:

• Recommended for You algorithm uses a hybrid

machine learning algorithm with content-based

ﬁltering and similarity among users to suggest

restaurants

• 30 minute delivery algorithm selects restaurants

that can fulﬁll the delivery timeline

• Appetizers under e 10 algorithm selects popular

restaurants that have numerous appetizers below

the set price limit.

FooDroid is a recommendation system devel-

oped at University of Zurich (Runo and Wattenhofer,

2011), created to provide a uniﬁed platform to man-

age booking of meals by students, having several

canteens with different menus available. Users can

browse the daily menus offered by the various can-

teens, select the meals according to their tastes and

evaluate them on behalf of the colleagues who will

later choose those based on the reviews made. This

recommendation system focuses mainly on:

• Detailed evaluation of the menus offered. In-

deed, since recommendation accuracy is mainly

affected by the (few) users providing the most rat-

ings, some menu choices could not reﬂect the real

taste of an individual. Therefore, a more personal-

ized recommendation system that operates on the

preferences of individuals (or small groups) is de-

sirable.

• Distance of the canteen from the customer’s loca-

tion, as it is assumed that you can only spend a

limited time for lunch. Thderefore, authors try to

obtain a trade off between the quality of the menus

and the relative proximity to the customer.

FooDroid aims at obtaining a good combination

between the quality of the menu offered by a can-

teen and users preferences, providing students with a

means of evaluating the meals consumed and recom-

mending menus based on these scores in the future.

Snap-n-eat (Zhang et al., 2015) is a mobile food

recognition system based on a deep learning ap-

proach. The system can recognize the food and es-

timate the calories and nutritional content of the food

automatically without any user intervention. To iden-

tify foods, the system allows the user to simply take

a picture of the plate of food. The system detects

the salient region, crops the image and subtracts the

background accordingly. Basically, the app identiﬁes

which segments of the image contain food and then

tries to ﬁgure out what type of food is present in each

segment. In addition, the system determines the por-

tion size which is then used to estimate the calories

and nutritional content of the food on the plate. The

system is capable of achieving automatic food detec-

tion and recognition in real life contexts containing

bulky backgrounds. When multiple items of food ap-

pear in an image, the system is able to identify them

and estimate their portions simultaneously.

3 THE CASE STUDY OF

WORKSITE CANTEEN

The FR system we propose here advises employees on

the choice of meal on a daily basis according to the

available menu; a mobile application (not discussed

here) allows to perform these and other actions, as

news management, meeting rooms booking, work-

place lighting control etc.

To set up the FR system, several information from

user’s proﬁle are considered together with those re-

lated to dishes, as its description, the list of ingredi-

ents and/or the contents of the nutritional levels; de-

pending on data available, various machine learning

techniques and models for similarity assessment be-

tween different data can be taken into account.

In particular, for what concern meals, we must

consider that the combinations of dishes on the days

are extremely varied and although the dishes are re-

peated several times during the analysis period, there

is not a high degree of repetition in the menus.

In addition, we classify canteen users into Fre-

quent Visitors (FV) or Occasional visitors (OV),

where a newcomer is ﬁrst classiﬁed as OV and pro-

moted to a FV after a given number of canteen access;

our FR system must address both categories.

3.1 Dataset

The dataset includes data collected during about 2.5

years, from August 2017 to March 2020, concerning

dishes consumed at the worksite, an innovation hub

Food Recommendation in a Worksite Canteen

119

Table 1: Employees distribution over companies.

CompanyID Meals Employees Average

meals

CompanyID Meals Employees Average

meals

0 21030 129 163.0 1 4679 48 97.5

2 7624 29 262.9 3 1593 264 6.0

4 1846 11 167,8 5 76 1 76.0

6 3671 10 367.1 7 6991 85 82.2

8 1206 11 109.6 9 303 3 101.0

10 175 4 43.8 11 2 2 1.0

12 27 1 27.0 13 3 2 1.5

14 72 4 18.0 15 4 1 4.0

16 2 1 2.0 18 2 1 2.0

19 239 33 7.2 20 32 4 8.0

21 13 1 13.0 22 3 1 3.0

Table 2: Dishes distribution.

Dish

category

Description Choices

1 Bread&pizza 8

2 Cold Cuts 17

3 First Course 158

4 Main Course 123

5 Salads 18

where workers of 22 different companies actually op-

erate.

The study of the employees food intake we focus

here is part of a more extensive project for life qual-

ity improvement of the employees at the worksite, in-

creasing their productivity through a better sense of

belonging (Carchiolo et al., 2019).

The number of meals consumed was about 50 000

(49 539), chosen by 264 employees. For a detailed

view on the number of employees of the companies

and the number of meals consumed, we refer the

reader to Table 1. The common canteen is acces-

sible by all the employees through a cross-platform

proprietary App, which provides the canteen manage-

ment with statistical information while retaining the

order and access history of each employee. Employ-

ees are identiﬁed with a ”employer code” that permits

us to know the company where they work, their ac-

cess to the canteen and the dishes consumed. Records

with missing values are pruned away from the data

set; this occurred for about 20% of meals. The dishes

proposed are organized in 5 categories that mimic the

italian mean of food. The total number of different

dishes offered by the canteen is 314.

We note that the multiplicity of proposals in the

different categories is not homogeneous and two of

them are way larger than others (see Table 2).

Each dish is described by:

• the ingredients

• the type of cooking

Figure 1: # of ingredient for dishes.

• the features of dietary restrictions

• the heat intake and nutrients present

The ingredients are obtained on the most common

Italian recipes. In the proposed dish 143 ingredients

are used and the number of main ingredients in each

dish ranges from 1 to 11, as depicted in Fig 1. In

ﬁg. 2 the distribution of ingredients in dishes is re-

ported. As shown, some ingredients are very com-

mon while others appear only occasionally in differ-

ent dishes. The most widely used ingredient is extra

virgin olive oil used for the preparation of 83% (in

ﬁg. 2 the red circle on the top) of the dishes; other

4 ingredients are very common in the dishes (they are

highlighted with the second red circle) and are present

in about 30% of the dishes. These ingredients reﬂect

the tradition of Italian cuisine.

We have also identiﬁed 5 different types of cook-

ing: fried, boiled, stew, grilled and baked. Each dish

can have a combination of these types of cooking

(even none of them). In ﬁg. 3 the distribution of types

of cooking in the dish is shown. In particular, the im-

pact of the single cooking method and the multiplicity

of that in a dish are illustrated.

As a restriction to the diet, those belonging to the

”gluten free” and ”vegetarian” categories are identi-

ﬁed as the characterization of the dishes. In ﬁg. 4

it is shown the distribution of veggie and gluten free

dishes.

COMPLEXIS 2021 - 6th International Conference on Complexity, Future Information Systems and Risk

120

Figure 2: Ingredients for dishes.

Figure 3: Cooking method for dishes.

Figure 4: Diet restriction.

3.2 Food Recommendation System

As already introduced, machine learning algorithms

used in recommendation systems are typically classi-

ﬁed into two main categories: content based and col-

laborative ﬁltering methods, although modern recom-

mendation systems combine both approaches. In our

solution we use a popularity based approach for the

Occasional Visitors, while we use an item–based col-

laborative ﬁlter for Frequent Visitors. For this reason

the data will be prepared differently in the two cases.

As previously discussed, the average number of

meals consumed in the canteen is 187, but only for

about 80% of them the dataset contains details of the

choice. An analysis of the dataset shows that the

distribution of accesses is not uniform at all. Fig. 5

shows that many employees access the canteen infre-

quently (gray bar), while others access a much higher

number of times. Fig. 5 shows in blue the variety in

the choice of dishes by each person.

Figure 5: # of meals and # of different choice of Complete

Data Set.

On the base of this analysis, we consider the value

of 100 meals already consumed at the canteen as the

threshold for identifying the FV. In the case of ”Fre-

quent Visitors” (FV), ﬁg. 6 shows the variety in the

choice of dishes by each person (blue bar), while gray

bar shows the number of times each person accesses

the canteen.

Figure 6: # of meals and # of different choice for FV.

In ﬁg 7 the number of times a dish has been chosen

by the FV is reported. As shown, there are dishes that

are much more common than others, meaning that

when using a polarity-based recommendation system

these will be the dishes offered.

3.2.1 Case 1: Frequent Visitors

To solve our problem either user–based or item–based

collaborative ﬁltering can be used. In the latter case,

the selection of the item (namely, the dish) is made

adopting a similarity based ﬁltering among dishes.

The goodness of the results of a recommendation sys-

tem is affected by the model used for coding the de-

Food Recommendation in a Worksite Canteen

121

Figure 7: Distribution of meal choice.

scription of the dishes and therefore to the algorithm

used to evaluate the similarity; several approaches can

be used:

• encoding the string of dish description D, by ap-

plying the one-hot encoder, to generate the array

A D describing the plate

• using a string SI obtained by concatenating the list

of ingredients included in the dish and applying

the one-hot encoder to generate the array A SI de-

scribing the plate

• using the product between array A SI with the ma-

trix containing the amount of each ingredient in

the plate

• adding the description of the dish with the amount

of nutrients.

Two dishes will be considered similar according

to the ﬁrst semantic model, which is the simplest to

implement. The differences arising from the choice

of the model are not analyzed in this paper.

In the case of a user–based collaborative ﬁltering,

a common solution is the use of some ranking. Our

Table 3: Test results.

FV FV + OV

RMSE CV RMSE CV

KNN 0.973 0.955 0.948 0.965

SVD 0.972 0.954 0.949 0.966

dataset does not include a ranking among elements,

but it is built using the history about the past choices

of dishes. In particular, a table is extracted from

the dataset in which each row contains the informa-

tion about the menu offered on a given day, the user

who made the reservation and his choices. To solve

our problem we made a test with both the classical

approaches of collaborative ﬁltering: memory–based

and a model–based. For each model, we perform two

experiments, one on data about FV’s and another on

the complete data set (FV+OV).

In the former case, we used a KNN (K-Nearest

Neighbours) with means model. This algorithm uses

the similarities between users and/or items as weights

to predict a rating for them. This similarity is com-

puted by using the Pearson correlation or cosine sim-

ilarity function.

The second approach used is based on the SVS

Model Based Collaborative Filtering. The Singular-

Value Decomposition, or SVD for short, is a matrix

decomposition method for reducing a matrix to its

constituent parts in order to make certain subsequent

matrix calculations simpler. It provides another way

to factorize a matrix, into singular vectors and singu-

lar values.

To ﬁnd the rating R that a user U would give to an

item I, the approach includes (1) ﬁnding users similar

to U who have rated the item I, and (2) calculating

the rating R based on the ratings of users found in

the previous step; when using KNN with means to

remove the bias we take into account the mean ratings

of each user.

To evaluate the results we calculate the RMSE

(Root Mean Squared Error), the most common met-

rics used to measure accuracy for continuous vari-

ables, and the Cross Validation (CV) of RSME with

KNN and SVD model; for each one we made a test

on the two above mentioned datasets: FV and (FV

+ OV). Table 3 summarizes the RSME and CV ob-

tained.

3.2.2 Case 2: Occasional Visitors

In deﬁning the recommendation systems, one of the

problems described is relating to cold starts, in partic-

ular when a new user is introduced into the data set.

Models deﬁned above do not make a correct predic-

tion on a user whose past choices are unknown, there-

COMPLEXIS 2021 - 6th International Conference on Complexity, Future Information Systems and Risk

122

Figure 8: Ranking for the most frequent visitor.

fore we adopt an alternative solution using a popu-

larity based recommender system that ranks products

based on their popularity (i.e. the rating count) at a

given moment. If a product is highly rated then it is

most likely to be ranked higher and hence will be rec-

ommended. As it is based on the products popularity,

the same set of products will be recommended for all

the users (no personalizing).

Data used to train the model contains User ID and

a Meal ID column. Each row represents an observed

interaction between the user and the element. Pairs

(user, meal) are stored with the model so that they

can later be excluded from recommendations if de-

sired. In addition, a Target column is added which

represents the Ranking of the choice. This ranking

is calculated as a function of the number of times the

person consumed a meal with respect to the total num-

ber of days that the meal was available and therefore

normalized in a range space.

Based on these assumptions, the initial data nec-

essary to create the model are different as the booking

history with the daily menu proposed and the choice

made is no longer required, rather it is only necessary

to obtain the sequence of choices made by each em-

ployee over time.

Results of such a model are at a very early stage

and the further development step envisaged is the re-

placement of the simple model based on popularity

with one that takes into account the group of people

with whom the new employee has already had con-

tacts because these with high probabilities will inﬂu-

ence the choices.

Since this is a popularity based recommender

model, we are getting similar result for all users. i.e.

the model is recommending same products for all the

users. Then we will choose meals recommended by

most popular persons (ﬁg 8 shows the results).

4 CONCLUSIONS AND FUTURE

WORK

In this paper we described a food recommendation

system for users of a workplace canteen. Exploiting a

dataset of past choices and using a content based and

a collaborative ﬁltering approach for canteen users,

the proposed system suggests users dishes chosen by

other similar users. First results for frequent as well

as occasional canteen visitors are encouraging though

further steps are need to validate the proposed ap-

proach, in particular:

• to leverage dishes complete features, as ingredi-

ents and macro nutrients

• to combine time series and machine learning for

prediction purposes, to help both canteen manager

for a better menu planning as well as users for a

better choice

• to test the proposed recommendation system in or-

der to validate it through users feedback

As future extensions, it is planned to use an iot based

approach (Loria et al., 2017) to collect other data,

such as movements and contacts between different

employees, useful for improving user proﬁling.

Food Recommendation in a Worksite Canteen

123

REFERENCES

Bobadilla, J., Ortega, F., Hernando, A., and Gutierrez, A.

(2013). Recommender systems survey. Knowledge-

Based Systems, 46:109 – 132.

Carchiolo, V., Grassia, M., Longheu, A., Malgeri, M., and

Mangioni, G. (2020). A network-based analysis to un-

derstand food-habits of a canteen clients. In Proceed-

ings of the 22nd International Conference on Informa-

tion Integration and Web-based Applications Services

iiWAS ’20.

Carchiolo, V., Grassia, M., Longheu, A., Malgeri, M., and

Mangioni, G. (2021). A network-based analysis of

a worksite canteen dataset. Big Data and Cognitive

Computing, 5(1).

Carchiolo, V., Longheu, A., Barbera, S., and Tommaso, P.

(2019). Personalized wellness in smart ofﬁces. In Pro-

ceedings of IADIS International Conference on Ap-

plied Computing 2019, pages 97–104.

Elsweiler, D., Trattner, C., and Harvey, M. (2017). Ex-

ploiting food choice biases for healthier recipe rec-

ommendation. In Proceedings of the 40th Interna-

tional ACM SIGIR Conference on Research and De-

velopment in Information Retrieval, SIGIR ’17, page

575–584, New York, NY, USA. Association for Com-

puting Machinery.

Freyne, J. and Berkovsky, S. (2010). Intelligent food plan-

ning: Personalized recipe recommendation. In Pro-

ceedings of the 15th International Conference on In-

telligent User Interfaces, IUI ’10, page 321–324, New

York, NY, USA. Association for Computing Machin-

ery.

Freyne, J., Berkovsky, S., and Smith, G. (2011). Recipe

recommendation: Accuracy and reasoning. In Pro-

ceedings of the 19th International Conference on User

Modeling, Adaption, and Personalization, UMAP’11,

page 99–110, Berlin, Heidelberg. Springer-Verlag.

Ge, M., Elahi, M., Ferna

andez-Tob

ıas, I., Ricci, F., and

Massimo, D. (2015). Using tags and latent factors in a

food recommender system. In Proceedings of the 5th

International Conference on Digital Health 2015, DH

’15, page 105–112, New York, NY, USA. Association

for Computing Machinery.

Harvey, M. and Elsweiler, D. (2015). Automated recom-

mendation of healthy, personalised meal plans. In Pro-

ceedings of the 9th ACM Conference on Recommender

Systems, RecSys ’15, page 327–328, New York, NY,

USA. Association for Computing Machinery.

Harvey, M. and Elsweiler, D. (2017). You are what you eat:

Learning user tastes for rating prediction. In Springer,

C., editor, Kurland O., Lewenstein M., Porat E. (eds)

String Processing and Information Retrieval. SPIRE

2013., Lecture Notes in Computer Science.

HBS-Digital-Initiative (2020). Caviar - premium bites at

your doorstep. last-access: 2020.

Iwendi, C., Khan, S., Anajemba, J. H., Bashir, A. K., and

Noor, F. (2020). Realizing an efﬁcient iomt-assisted

patient diet recommendation system through machine

learning model. IEEE Access, 8:28462–28474.

Jiang, H., Wang, W., Liu, M., Nie, L., Duan, L.-Y., and

Xu, C. (2019). Market2dish: A health-aware food

recommendation system. In Proceedings of the 27th

ACM International Conference on Multimedia, MM

’19, page 2188–2190, New York, NY, USA. Associa-

tion for Computing Machinery.

Loria, M. P., Toja, M., Carchiolo, V., and Malgeri, M.

(2017). An efﬁcient real-time architecture for col-

lecting iot data. In 2017 Federated Conference on

Computer Science and Information Systems (FedC-

SIS), pages 1157–1166.

Merler, M., Wu, H., Uceda-Sosa, R., Nguyen, Q.-B., and

Smith, J. R. (2016). Snap, eat, repeat: A food recog-

nition engine for dietary logging. In Proceedings of

the 2nd International Workshop on Multimedia As-

sisted Dietary Management, MADiMa ’16, page 31–

40, New York, NY, USA. Association for Computing

Machinery.

Min, W., Jiang, S., Liu, L., Rui, Y., and Jain, R. (2019).

A survey on food computing. ACM Comput. Surv.,

52(5).

Mohamed, M., Khafagy, M., and Ibrahim, M. (2019). Rec-

ommender systems challenges and solutions survey.

Nations, U. (2015). Transforming our world: the 2030

agenda for sustainable development. Technical report.

https://sustainabledevelopment.un.org/resourcelibrary.

Runo, M. and Wattenhofer, R. (2011). Foodroid: A food

recommendation app for university canteens. Master’s

thesis, ETH Zurich, Zurich, Switzerland.

Torreggiani, S., Mangioni, G., Puma, M. J., and Fagiolo,

G. (2018). Identifying the community structure of the

food-trade international multi-network. Environmen-

tal Research Letters, 13(5):054026.

Yang, L., Hsieh, C.-K., Yang, H., Pollak, J. P., Dell, N.,

Belongie, S., Cole, C., and Estrin, D. (2017). Yum-

me: A personalized nutrient-based meal recommender

system. ACM Trans. Inf. Syst., 36(1).

Zhang, M., Tian, G., Zhang, Y., and Duan, P. (2021). Rein-

forcement learning for logic recipe generation: Bridg-

ing gaps from images to plans. IEEE Transactions on

Multimedia, pages 1–1.

Zhang, W., Yu, Q., Siddiquie, B., Divakaran, A., and Sawh-

ney, H. (2015). Snap-n-eat: Food recognition and nu-

trition estimation on a smartphone. Journal of dia-

betes science and technology, 9(3).

Zhang, X., Luo, X., and Chen, B. (2020). Multi-view visual

bayesian personalized ranking for restaurant recom-

mendation. Applied Intelligence.

COMPLEXIS 2021 - 6th International Conference on Complexity, Future Information Systems and Risk

124