Multiobjective recommendation for sustainable production systems

Abstract

We present a recommendation system to help rebuild sustainable production systems. Our multi-objective system synergizes the public and private actors of a territory. From know-how proximities in the Product Space, we suggest productive jumps for companies in a territory that consider the expectations of companies not only in terms of diversification but also in terms of the expectations of local authorities who are anxious to build sustainable production systems. We formalize a multi-stakeholder recommendation that is applied to the sustainability of a territorial economy and we propose the following new objectives to consider: (i) Economic growth, based on the concept of territorial economic complexity ; (ii) Productive resilience, defined rigorously from the theory of dynamic systems; (iii) Food security and more generally basic necessities from an original approach based on Maslow's hierarchy of needs; (iv) The need to develop greener productions that respect the environment. The recommendation system that we propose incorporates territorial policy as a weighting of objectives. This "configuration" acts directly on the system to influence the recommended productive jumps. Each objective is defined to be computed directly from open data available for most countries without requiring external data.

Full text

Multiobjective recommendation for sustainable production systems∗
ARNAULT PACHOT, ADÉLAÏDE ALBOUY-KISSI, BENJAMIN ALBOUY-KISSI, and FRÉDÉRIC CHAUSSE,
Uni-
versité Clermont-Auvergne, CNRS, SIGMA Clermont, Institut Pascal, France
We present a recommendation system to help rebuild sustainable production
systems. Our multi-objective system synergizes the public and private actors
of a territory. From know-how proximities in the Product Space, we suggest
productive jumps for companies in a territory that consider the expectations
of companies not only in terms of diversication but also in terms of the
expectations of local authorities who are anxious to build sustainable pro-
duction systems. We formalize a multi-stakeholder recommendation that
is applied to the sustainability of a territorial economy and we propose the
following new objectives to consider:
(i)
Economic growth, based on the concept of territorial economic com-
plexity;
(ii)
Productive resilience, dened rigorously from the theory of dynamic
systems;
(iii)
Food security and more generally basic necessities from an original
approach based on Maslow’s hierarchy of needs;
(iv)
The need to develop greener productions that respect the environ-
ment.
The recommendation system that we propose incorporates territorial policy
as a weighting of objectives. This "conguration" acts directly on the system
to inuence the recommended productive jumps. Each objective is dened
to be computed directly from open data available for most countries without
requiring external data.
CCS Concepts:
•Applied computing →Supply chain management
;
•
Information systems →Recommender systems.
Additional Key Words and Phrases: Multi-Objective Recommender Systems,
Supply-chain resilience, Sustainable production system
1 INTRODUCTION
The Covid 19 crisis has shown the fragility of our European produc-
tion systems. Years of externalizations have damaged our productive
capacity. However, a general awareness has emerged following the
crisis, and the public and private sector are now ready to collabo-
rate to rebuild a sustainable production system. Financial support
programs have been deployed to help companies relocate their pro-
duction or reinforce their existing activities.
At the same time, companies have understood the importance of
securing their supplies through local production units. This oers
new business opportunities to suppliers to develop their production
towards new products to overcome shortages. To support this eort,
we imagined a recommender system whose goal is to suggest the de-
velopment of new products to companies to ensure their commercial
development, while taking into account territorial policies.
We will start by presenting the area of recommendation. Then,
after describing the previous work in relation to the eld of industry,
∗
Copyright 2021 for this paper by its authors. Use permitted under Creative Commons
License Attribution 4.0 International (CC BY 4.0).
Presented at the MORS workshop held in conjunction with the 15th ACM Conference
on Recommender Systems (RecSys), 2021, in Amsterdam, Netherlands.
Authors’ address: Arnault Pachot, arnault.pachot@etu.uca.fr; Adélaïde Albouy-Kissi,
adelaide.kissi@uca.fr; Benjamin Albouy-Kissi, benjamin.albouy@uca.fr; Frédéric
Chausse, frederic.chausse@uca.fr, Université Clermont-Auvergne, CNRS, SIGMA Cler-
mont, Institut Pascal, , Clermont-Ferrand, France, F-63000.
we will introduce the public and private stakeholders. We will then
detail how our multi-objective recommendation system works and
we will present a rst experimentation on a French territory.
2 PREVIOUS WORKS
Recommendation systems are very successful in many areas. There
are two types of these systems: content-based and collaborative
ltering[
6
,
7
]. Several recommendation systems are derived from the
two main types, and it is usual to combine them before generating
the list of recommended objects[10].
2.1 Multi-objective recommendation
Traditionally, recommendation systems are oriented towards the
end-user and seek to optimize a single cost function. However,
recommendation methods can be multi-objective when they aim to
optimize several objectives. For example, by integrating diversity
and novelty in the proposed products[
4
,
46
,
48
,
55
]. Other uses
concern the consideration of price in recommended products[
15
,
30
].
Recommendation systems that seek to improve the fairness of results
are also multi-objective recommendation systems[
11
,
40
,
56
]. The
diculty is to take into account each objective without signicantly
degrading the accuracy on the main objective.
Several approaches to this problem exist. The rst approach draws
its foundations from multi-objective optimization and seeks to opti-
mize all of the objectives at the same time. This approach is based
on Pareto concepts and the associated algorithms are of the evolu-
tionary type[13, 17, 35, 41, 55, 57].
The second approach considers multi-objectives as a hybridiza-
tion of methods in which the combination of results can be done in
cascade[
31
,
36
]: an objective renes the result of a previous objective
(re-ranking), or mixing[
46
]. The multi-objective recommendation
is then considered as a weighted hybridization of mono-objective
functions.
2.2 Multistakeholder recommendation
Multistakeholder recommendation systems[
2
,
3
,
5
,
12
] are derived
from multi-sided platforms[
16
,
47
] and reciprocal recommendation
systems. The latter require us to take each stakeholder into account
independently because their strategies, and therefore their objec-
tives, are dierent.
2.3 Recommendation in the field of industry
Pachot et al
. [43]
have developed a recommendation system to
recommend collaborative synergies between companies in the same
territory based on the semantic analysis of product nomenclatures
to nd the productive link that exists (for example) between seed,
wheat, our and bread. A distribution of products in a vector space
allows us to make recommendations.
The recommendations can be similar to client-supplier or co-
production relationships. To improve the industrial resilience of a
, Vol. 1, No. 1, Article . Publication date: September 2021.

2•Arnault Pachot, Adélaïde Albouy-Kissi, Benjamin Albouy-Kissi, and Frédéric Chausse
territory, it is appropriate to develop distributed manufacturing by
encouraging the companies of a territory to work with each other.
The recommendation system integrates an alternative operation
when no potential supplier is present on a territory, which suggests
that suppliers are able to make "productive jumps" to produce the
required goods. These productive jumps are made possible by the
productive relationship between the classes of products. This relies
on the data of productive proximity from the Product Space[27].
The modeling of the companies’ productions is based on a sta-
tistical analysis of the productions associated with their economic
activity code. A rst experiment was carried out on French compa-
nies.
We also nd recommendation systems dedicated to the supply
chain, specically for distribution[
14
,
29
] or to promote the use of
waste in the context of industrial symbiosis[54].
To our knowledge, no study has aimed at the construction of a
multi-objective recommendation system in the eld of industrial
production, and in particular with the aim of favoring the construc-
tion of sustainable production systems.
3 DESCRIPTION OF STAKEHOLDERS
3.1 Companies
Companies produce manufactured goods or raw materials. Their
production units are located on a territory and carry out an economic
activity. Firms collaborate with each other within a territory, or
import goods or raw materials from other territories or countries.
Companies follow a commercial strategy that not only encour-
ages them to develop their commercial portfolio by seeking new
customers but also encourages them to diversify their production by
favoring the production of goods that give them a better competitive
advantage. At the same time, they seek to secure their supplies by
diversifying their suppliers and giving preference to local suppliers.
For several years, companies have also been encouraged to improve
their social and environmental impact.
3.2 Local authorities
Building a sustainable ecosystem requires active collaboration be-
tween the private and public sectors. We would like to integrate
local authorities into our recommendation system, which through
their nancial aid, taxation or thanks to their teams on the ground
have a certain number of levers to help build such ecosystems.
The territorial policy that is associated with local authorities is
dened by several objectives related to economic growth, food secu-
rity, industrial resilience, and environmental aspects. We consider
territorial policy as a "conguration" of the recommendation system
in which local authorities indicate their priorities on each of the
objectives.
4 FORMALIZATION
We design a recommendation system for the companies of a terri-
tory, whose object is the recommendation of new products to be
developed. These production units have a know-how that oers
them the possibility to make "productive jumps"; that is, to move
from the production of one kind of product to another when the
"proximity of know-how" between the two products is relatively
strong.
There is naturally a propensity for companies to adapt their of-
fer to seize new commercial opportunities, but we propose to set
up a recommendation system that also takes into account territo-
rial policy. As mentioned earlier, the territorial policy consists in
weighing the dierent objectives of the referral system. The public
authorities thus have the possibility of inuencing the functioning
of the system by choosing the objectives that are most important to
them. Let
{𝑎1, 𝑎2, . . . , 𝑎𝑛}
be the list of algorithms associated with
each objective. The territorial policy
P
on the territory
𝜏
is dened
as follows:
P(𝜏)={𝑤𝑎0, 𝑤𝑎1, . . . ,𝑤 𝑎𝑛}(1)
We start by listing the achievable production jumps for a pro-
duction unit that correspond to the rst objective for companies:
diversifying their production. To do this, we rst calculate the cur-
rent production associated with each production unit. In this task,
we rely on a correspondence table that makes the link between the
economic activity code of the production unit and the associated
production (product codes in the HS nomenclature).
We then directly use the Product Space developed by the Growth
Lab of Harvard University[
24
] to identify the opportunities for
productive jumps for each of the products that the production unit
manufactures. The Product Space is a graph of products in which
each node corresponds to a product class (from the HS classication)
and the weighting of the edges corresponds to the proximity of
know-how between two product classes. For a given production unit
𝑢
, we obtain a list of new products
A
:
{𝑥0, 𝑥1, . . . , 𝑥 𝑛}
, which are
ranked in descending order with respect to the level of productive
relatedness. We choose an additional objective for companies that
consists in increasing the competitive advantage and four objectives
for territorial authorities that make sense with the development of
sustainable systems[
42
]. They take into account economic aspects,
resilience, security of basic goods and environment:
•
Economic growth: we integrate the objective of developing
the economic growth of the territory. This is a wealth cre-
ator, and is essential to ensure economic prosperity and job
creation. To identify the products that are the most eective
in creating economic growth, we will take into account their
level of economic complexity, which is a measure that has
been shown to be highly correlated.
•
Productive resilience: we integrate an objective to improve
the level of resilience of a territory. Given the fragility of our
production systems, which have been damaged by various
economic or health crises, we absolutely must build more
robust production systems. We are going to integrate a theo-
retical measure of resilience of a territory.
•
Securing basic necessities: we also want to give the possibility
to favor the basic necessities (e.g., food and pharmaceuticals)
over other products. To do so, we have followed an origi-
nal approach inspired by Maslow’s hierarchy to distinguish
"vital" products.
•
Green production: nally, we take into account the environ-
mental dimension of production, aware of the importance of
, Vol. 1, No. 1, Article . Publication date: September 2021.

Multiobjective recommendation for sustainable production systems •3
Table 1. Stakeholders in the recommender system
Stakeholders Strategies Objectives Function
Companies Business strategy Diversication 𝑎1
Competitive advantage 𝑎2
Local authorities Territorial policy Economic growth 𝑎3
Productive resilience 𝑎4
Securing basic necessities 𝑎5
Green production 𝑎6
repositioning production systems towards the production of
greener goods.
We present below the technical details of the measurement of
each of these objectives. We perform a weighted hybridization[
46
]
from the dierent objectives to re-rank the list A.
For each production unit
𝑢
we compute the scores
{ˆ
𝑝𝑎1(𝑥𝑖|𝑢),
ˆ
𝑝𝑎2(𝑥𝑖|𝑢), . . . , ˆ
𝑝𝑎𝑛(𝑥𝑖|𝑢)}
of each product
𝑥𝑖∈ A
for each algorithm
{𝑎1, 𝑎2, . . . , 𝑎𝑛}
. All scores must be normalized. Then we perform a
weighted sum of each score to obtain a nal score for each product:
ˆ
𝑝(𝑥𝑖|𝑢)=
𝑛
Õ
𝑗=1
ˆ
𝑝𝑎𝑗(𝑥𝑖|𝑢) × 𝑤𝑎𝑗(2)
4.1 Objective 1: Diversify production
Our system makes a recommendation of productive jumps, which
are repositioning or industrial diversication opportunities for com-
panies. As presented in Pachot et al
. [43]
, the Product Space de-
veloped by Hausmann and Klinger
[25]
provides a model to make
recommendations.
To make the correspondence between a production unit and the
products it manufactures, we use a correspondence table
1
. The pro-
ductive proximity between product classes is based on the study of
country co-exports. From a large-scale analysis of the types of prod-
ucts exported by country, Hidalgo et al
. [27]
computed the proximity
of productive know-how (called "productive kinship") between each
type of product and construct a graph of the productive space.
The measurement of the productive proximity between each
product is done by looking for the percentage of times that product
𝑝1is co-exported with product 𝑝2:
𝜙𝑝1,𝑝2=𝑚𝑖𝑛 Í𝑐𝑀𝑐𝑝1𝑀𝑐𝑝 2
Í𝑐𝑀𝑐𝑝1Í𝑐𝑀𝑐𝑝 1𝑀𝑐𝑝2
Í𝑐𝑀𝑐𝑝2(3)
We consider that a product
𝑝
is exported by a country
𝑐
when it
grants the country a revealed competitive advantage (RCA) accord-
ing to the formula of Balassa
[8]
. Let
𝑋𝑐𝑝
be the exports of product
𝑝
by country
𝑐
, then the revealed competitive advantage that country
𝑐has for product 𝑝can be expressed as a function of exports:
RCAcp =
𝑋𝑐𝑝
Í𝑐𝑋𝑐𝑝
/Í𝑝𝑋𝑐𝑝
Í𝑐,𝑝 𝐶𝑐𝑝
(4)
1
There are several correspondence tables according to the nomenclatures used
for the classes of economic activities. For example, for the USA, ISIC-HS
(https://unstats.un.org/unsd/classications/Econ) or NAF-CPF, equivalent to NACE-
CPA for Europe (https://www.insee.fr/fr/information/2399243). See Pachot et al
. [43]
for details
We consider that a country
𝑐
exports a product
𝑝
if
𝑅𝐶𝐴𝑐𝑝
is
greater than 1.
𝑀𝑐𝑝 =1𝑖 𝑓 RCAcp ≥1;
0𝑒𝑙𝑠𝑒 (5)
From the data of productive proximity available in open data[
53
],
we build a function
𝑎1(𝑢)
that for each production unit
𝑢
will asso-
ciate a list
A
of products and their associated scores of productive
proximity.
4.2 Objective 2: Increase the competitive advantage
Now we need to consider the commercial interest for each rm.
Some products are more advantageous for a production unit than
others and we have the
RCA
formula to allow us to rank the pro-
ductive opportunities according to the competitive advantage that
they would grant to the production unit.
We use a modied version of
RCA
applied to the products of
a sub-national territory. We compare the share of an activity in
a territory with the share of that activity on a global scale. This
prevents the more developed regions of a country from appearing
to have a comparative advantage in each product[9, 45]:
RCA𝑙𝑜𝑐𝑎𝑙
𝑐𝑝 =
𝑋𝑙𝑜𝑐 𝑎𝑙
𝑐𝑝 /𝑋𝑙𝑜𝑐 𝑎𝑙
𝑐
𝑋𝑤𝑜𝑟𝑙𝑑
𝑝/𝑋𝑤𝑜𝑟𝑙 𝑑 (6)
We dene a function
𝑎2(𝑢, 𝜏 )
, which for each production unit
𝑢
of a territory
𝜏
will associate a list of products
∈ A
and their
associated score of RCA.
4.3 Objective 3: Improve economic performance
Several studies have conrmed the strong relationship between a
country’s economic complexity and its growth rate: regions spe-
cializing in the manufacture of more complex products experience
faster economic growth[
18
,
27
]. Therefore, we choose to use the
economic complexity indicator (ECI) as a target to improve the
growth of a territory.
The calculation of economic complexity presented by Hausmann
et al
. [24]
is based on two measures: productive diversity and ubiq-
uity. Diversity illustrates the variety of dierent products exported
by a country. Ubiquity is an indication of the number of countries
that export the same product. Let
𝑀
be a matrix of products ex-
ported by country, such that
𝑀𝑐𝑝 =
1if country
𝑐
exports product
𝑝:
Diversity :𝑘𝑐,0=Í𝑝𝑀𝑐𝑝
Ubiquity :𝑘𝑝,0=Í𝑐𝑀𝑐𝑝 (7)
, Vol. 1, No. 1, Article . Publication date: September 2021.

4•Arnault Pachot, Adélaïde Albouy-Kissi, Benjamin Albouy-Kissi, and Frédéric Chausse
From the measures of diversity
𝑘𝑐,0
and ubiquity
𝑘𝑝,0
, we can
recursively dene the variables
𝑘𝑐,𝑁
and
𝑘𝑝,𝑁
corresponding to the
average ubiquity and diversity of products exported by a country
𝑐
.
𝑘𝑐,𝑁 =1
𝑘𝑐,0Í𝑝𝑀𝑐𝑝 . 𝑘𝑝 ,𝑁 −1
𝑘𝑝,𝑁 =1
𝑘𝑝,0Í𝑐𝑀𝑐𝑝 . 𝑘𝑐,𝑁 −1(8)
We wish to express
𝑘𝑐,𝑁
as a function of
𝑘𝑐,0
. To do this, we replace
𝑘𝑝,𝑁 −1by 1
𝑘𝑝,0Í𝑐𝑀𝑐𝑝 . 𝑘𝑐,𝑁 −2and then simplify the equation:
𝑘𝑐,𝑁 =1
𝑘𝑐,0Í𝑝𝑀𝑐𝑝 1
𝑘𝑝,0Í𝑐′𝑀𝑐′𝑝. 𝑘𝑐′,𝑁 −2
=Í𝑐′𝑘𝑐′,𝑁 −2Í𝑝
𝑀𝑐′𝑝𝑀𝑐𝑝
𝑘𝑐,0𝑘𝑝,0
(9)
We consider the matrix
∼
𝑀𝑐,𝑐′
as a weighted (and normalized)
diversication similarity matrix. This matrix reects the extent
to which the types of products exported from two countries are
similar[38]:
∼
𝑀𝑐,𝑐′≡Õ
𝑝
𝑀𝑐′𝑝𝑀𝑐𝑝
𝑘𝑐,0𝑘𝑝,0
(10)
Let us rewrite the equation:
𝑘𝑐,𝑁 =Õ
𝑐′
∼
𝑀𝑐,𝑐′. 𝑘𝑐′,𝑁 −2(11)
We note that
𝑘𝑐,𝑁 =𝑘𝑐, 𝑁 −2=
1satises this equation when the
eigenvector of
∼
𝑀𝑐,𝑐′
is associated with the largest eigenvalue. Since
this eigenvector is a vector composed only of 1, it does not contain
any information. Therefore, it is better to look at the eigenvector
that is associated with the second largest eigenvalue. This is the
eigenvector that captures the most variance in the system and is
therefore a relevant measure of economic complexity. We denote
®
𝐾𝑖
as the i-th eigenvector of ∼
𝑀𝑐,𝑐′, ordered in a decreasing order[22]:
∼
𝑀𝑐,𝑐′.®
𝐾2=𝜆2®
𝐾2(12)
The economic complexity index
𝐸𝐶𝐼
[
26
] is obtained by normal-
izing the eigenvector of
∼
𝑀𝑐,𝑐′
associated with its second largest
eigenvalue.
<®
𝐾2>
corresponds to the mean of
®
𝐾2
and
𝑠𝑡𝑑 𝑒𝑣 (®
𝐾2)
to its standard deviation.
ECI𝑐=
®
𝐾2−<®
𝐾2>
𝑠𝑡𝑑 𝑒𝑣 (®
𝐾2)(13)
In the same way, the complexity of a product (PCI) is dened by
the following formula, with
®
𝑄
the eigenvector of
∼
𝑀𝑝,𝑝′
constructed
on the same principle as
∼
𝑀𝑐,𝑐′
, but exchanging the
𝑐
countries with
the 𝑝products:
PCI𝑝=
®
𝑄2−<®
𝑄2>
𝑠𝑡𝑑 𝑒𝑣 (®
𝑄2)(14)
We now wish to calculate the economic complexity of a sub-
national territory. We must use a modied version of the equation
13, which combines PCIs calculated using international trade data
with local data:
ECI𝑙 𝑜𝑐𝑎𝑙
𝑐=
1
𝑀𝑐Õ
𝑝
𝑀𝑙𝑜𝑐 𝑎𝑙
𝑐𝑝 PCI 𝑝(15)
Where
𝑀𝑙𝑜𝑐 𝑎𝑙
𝑐𝑝
is calculated in the same way as equation 5 but
using RCA𝑙𝑜𝑐𝑎𝑙
𝑐𝑝 instead of RCA𝑐𝑝 :
𝑀𝑙𝑜𝑐 𝑎𝑙
𝑐𝑝 =1𝑠𝑖 RCA𝑙𝑜 𝑐𝑎𝑙
𝑐𝑝 ≥1;
0𝑒𝑙𝑠𝑒 (16)
A pre-calculated table with the complexities associated with each
Harmonized System (HS) product class is available in open data
2
.
We choose to retain the values for the latest available year (i.e.,
2019).
We dene a function
𝑎3(𝑢)
, which for each production unit
𝑢
will
associate a list of products
∈ A
and their associated score of
PCI
ranked in decreasing order.
4.4 Objective 4: Improve the resilience of the production
system
Resilience is dened as the ability to recover quickly after a disrup-
tive shock. For a production system, this corresponds to the ability
of a system to quickly recover its production level or a higher level.
To measure the resilience indicator of a territorial production sys-
tem, we start from an approach derived from the theory of dynamic
systems[
49
–
51
]. In particular, the studies of Kharrazi
[32]
, Kharrazi
et al
. [33
,
34]
focus on the denition of a theoretical resilience in-
dicator built from the analysis of imports, while the exports of a
territory are of particular interest to us. The theoretical resilience
of a dynamic system is proposed based on two measures: eciency
and redundancy. Kharrazi et al
. [34]
have conducted a study on
the behavior of the production systems of countries between 1996
and 2012, including the economic crisis of 2009, conrming the
relevance of the theoretical indicator of resilience. The measure of
a territory’s eciency (also called ascendancy) can be considered
as the degree of articulation or constraint of ows in a production
system[
34
]. The more specialized a system is, the more optimized
its connections are, the more ecient it is, and the less resilient it
is. The theoretical measure of eciency is as follows:
Eciency =Õ
𝑖, 𝑗
𝑇𝑖𝑗
𝑇..
log 𝑇𝑖𝑗 𝑇..
𝑇𝑖.𝑇.𝑗
(17)
Where
𝑇𝑖𝑗
is a product export value from country
𝑖
to country
𝑗
,
𝑇𝑖. =Í𝑗𝑇𝑖𝑗
is the total exports leaving country
𝑖
,
𝑇.𝑗 =Í𝑖𝑇𝑖 𝑗
is the
total imports entering country
𝑗
and
𝑇.. =Í𝑖 𝑗 𝑇𝑖 𝑗
is the sum of all
exports in the system [21].
Conversely, a redundant system has many connections, and will
therefore be "more exible in re-rooting its ows and maintaining
critical functions"[
32
]. Redundancy can be dened as the "degree
of freedom or overhead of ows in a network"[
32
]. It is measured
from the conditional entropy:
Redundancy =−Õ
𝑖, 𝑗
𝑇𝑖𝑗
𝑇..
log
𝑇2
𝑖 𝑗
𝑇𝑖.𝑇.𝑗
(18)
2https://atlas.cid.harvard.edu/rankings/product
, Vol. 1, No. 1, Article . Publication date: September 2021.

Multiobjective recommendation for sustainable production systems •5
From the eciency and redundancy measures of a system, we
can then measure the theoretical resilience level:
𝛼=Eciency/(Eciency +Redundancy)
Resilience =−𝛼log(𝛼)(19)
We seek to recommend new products to be developed to com-
panies in a territory that can help to improve the resilience score.
We compute the contribution of a each product
𝑥
to the resilience.
We dene a function
𝑎4(𝑢)
which for each production unit
𝑢
will
associate a list of products
∈ A
and their associated score of their
contribution to the resilience of the territory.
4.5 Objective 5: Secure the production of essential goods
We consider that products can be ordered according to their contri-
bution to the needs of individuals. To do so, we propose an approach
based on the hierarchy of needs of Maslow [37].
Genkova
[20]
provides a table of correspondence between the
categories of needs of Maslow’s pyramid and the categories of prod-
ucts in the CPC nomenclature. All of the indirect products that are
necessary for the production of the goods of each category are also
associated.
We weight each product inversely to its corresponding level in
Maslow’s pyramid and we obtain a function
𝑎5(𝑢)
, which for each
production unit
𝑢
will associate a list of products
∈ A
and their
associated score of their contribution to the needs.
4.6
Objective 6: Promote the production of environmental
products
We want the recommendation system to take the environmental
impact of products into account. As a priority, we propose the pro-
ductive jumps towards products with a lower environmental impact.
Several studies have been carried out to integrate this dimension
into the Product Space [19, 23, 28, 39, 44].
Initiatives exist to list green products[
1
,
19
]. We retain the list
provided by APEC
3
, which is dened as products "that directly and
positively contribute to green growth and sustainable development
objectives".
Let
G
be the list of green products, we dene a function
𝑎5(𝑢)
,
which for each production unit
𝑢
will associate a list of products
𝑥∈ A
and an associated score
𝑠
such that if
𝑥∈ G
then
𝑠=
1else
𝑠=0.
5 EXPERIMENTATION
We tested the recommender system using open data on French com-
panies: production units, import and export amounts by French de-
partment and by product class. We made available the pre-calculated
rankings associated with each objective, as well as a rst experi-
mentation on a French department4.
3
APEC List of Environmental Goods: https://www.apec.org/meeting-papers/leaders-
declarations/2012/2012_aelm/2012_aelm_annexc.aspx
4
https://github.com/apachot/Multiobjective-recommendation- for-sustainable-
production-systems
5.1 Datasets
We relied on the SIRENE
5
dataset that is available history of French
production units since 1973. It provides information for every com-
pany, relating them to their connected production unit, their NACE
economic sector, their workforce group, and their postal address.
We choose the HS nomenclature limited to 4 digits as the reference
nomenclature for the products. By resorting to a combination of
correspondence tables between activities and products (NACE
→
CPA→HS)6, we associate each NACE class with HS classes.
We use datasets from global trade[
52
], as well as local datasets
from each French department
7
. These datasets provide us for a given
territory or country, the amount of exports of each product class,
for each country or french department. We use the year 2019 and
convert the French data (French CPF nomenclature) into the HS
nomenclature.
5.2 Recommender System
We retrieve the list of production units on a territory. From their
activity code, our system is able to determine which product classes
are manufactured by this production unit. We then use a proximity
table between the product classes to determine which products are
the closest in the sense of know-how. These products represent the
potential production jumps.
At each productive jump we compute a global score from the
weighted average of 6 pre-computed rankings, associated to the 6
objectives of our recommendation system. We then propose a list of 5
classes of products whose productive jump has obtained the highest
score. You will nd in table 2 an example of recommendation for a
production unit located in Haute-Loire in France, that manufactures
parts and accessories for motor vehicles.
6 CONCLUSION
The expectations of the dierent actors in a territory often come
up against the complexity of the production systems. Stakeholders
whose strategies are sometimes opposed can nd a solution in a
recommendation system that takes their objectives into account.
We explored the elds of recommendation, information theory and
economics to nd objectives that can be integrated, and we formal-
ized a rst version of a multi-objective recommendation system.
We will continue our work by validating the system on a territory.
We also consider a Pareto-ecient hybridization to guide the local
authorities in setting the weights of the system.
REFERENCES
[1]
2014. The Stringency of Environmental Regulations and Trade in Environmental
Goods. OECD Trade and Environment Working Papers 2014/03. https://doi.org/
10.1787/5jxrjn7xsnmq-en Series: OECD Trade and Environment Working Papers
Volume: 2014/03.
5https://www.sirene.fr/sirene/public/static/acces-donnees
6
We parse the NAF to CPF correspondence document (https://www.insee.fr/fr/
statistiques/chier/2399243/Nomenclatures_NAF_et_CPF_Reedition_2020.pdf) to ob-
tain a NACE 2 to CPF 2.1 CSV le. NACE is obtained from NAF code by remov-
ing the last letter. CPF and CPA are identical in version 2.1. Matching between
CPA 2.1 and HS 2017 is done using the CPA 2.1 to NC 2017 correspondence ta-
ble (https://ec.europa.eu/eurostat/ramon/relations/index.cfm?TargetUrl=LST_REL) be-
cause HS code equivalent to the sixth rst digits of the related NC code.
7
https://www.data.gouv.fr/fr/datasets/statistiques-regionales-et-departementales-
du-commerce- exterieur/
, Vol. 1, No. 1, Article . Publication date: September 2021.

6•Arnault Pachot, Adélaïde Albouy-Kissi, Benjamin Albouy-Kissi, and Frédéric Chausse
Table 2. Example of recommendations for a production unit located in Haute-Loire in France (NACE activity code 29.32: Manufacture of other parts and
accessories for motor vehicles).
Code Description Score 𝑎1𝑎2𝑎3𝑎4𝑎5𝑎6
HS8406 Turbines; steam and other vapour turbines 3.76 0.61 0.001 0.74 0.41 1 1
HS8514
Industrial or laboratory electric furnaces and ovens (in-
cluding those functioning by induction or dielectric
loss); other industrial or laboratory equipment for the
heat treatment of materials by induction or dielectric
loss
3.71 0.67 3.7e-06 0.89 0.16 1 1
HS8503
Electric motors and generators; parts suitable for use
solely or principally with the machines of heading no.
8501 or 8502
3.69 0.68 1.37e-05 0.70 0.31 1 1
HS8419
Machinery, plant (not domestic), or laboratory equip-
ment; electrically heated or not, (excluding items in
85.14) for the treatment of materials by a process in-
volving change of temperature; including instantaneous
or non electric storage water heaters
3.67 0.72 0.00024 0.78 0.17 1 1
HS8417
Furnaces and ovens; industrial or laboratory, including
incinerators, non-electric
3.54 0.58 1.8e-05 0.74 0.21 1 1
[2]
Himan Abdollahpouri, Gediminas Adomavicius, Robin Burke, Ido Guy, Dietmar
Jannach, Toshihiro Kamishima, Jan Krasnodebski, and Luiz Pizzato. 2020. Multi-
stakeholder recommendation: Survey and research directions. User Modeling and
User-Adapted Interaction 30, 1 (March 2020), 127–158. https://doi.org/10.1007/
s11257-019- 09256-1
[3]
Himan Abdollahpouri and Robin Burke. 2019. Multi-stakeholder Recommendation
and its Connection to Multi-sided Fairness. arXiv:1907.13158 [cs] (July 2019).
http://arxiv.org/abs/1907.13158 arXiv: 1907.13158.
[4]
Himan Abdollahpouri, Robin Burke, and Bamshad Mobasher. 2017. Control-
ling Popularity Bias in Learning-to-Rank Recommendation. In Proceedings of the
Eleventh ACM Conference on Recommender Systems. ACM, Como Italy, 42–46.
https://doi.org/10.1145/3109859.3109912
[5]
Himan Abdollahpouri, Robin Burke, and Bamshad Mobasher. 2017. Recommender
Systems as Multistakeholder Environments. In Proceedings of the 25th Conference
on User Modeling, Adaptation and Personalization. ACM, Bratislava Slovakia, 347–
348. https://doi.org/10.1145/3079628.3079657
[6]
G. Adomavicius and A. Tuzhilin. 2005. Toward the next generation of recom-
mender systems: a survey of the state-of-the-art and possible extensions. IEEE
Transactions on Knowledge and Data Engineering 17, 6 (June 2005), 734–749.
https://doi.org/10.1109/TKDE.2005.99
[7]
Charu C. Aggarwal. 2016. Recommender Systems. Springer International Publish-
ing, Cham. https://doi.org/10.1007/978-3- 319-29659-3
[8]
Bela Balassa. 1986. Comparative Advantage in Manufactured Goods: A Reap-
praisal. The Review of Economics and Statistics 68, 2 (May 1986), 315. https:
//doi.org/10.2307/1925512
[9]
Santiago Perez Balsalobre, Carlos Llano Verduras, and Jorge Diaz-Lanchas. 2019.
Measuring subnational economic complexity: An application with Spanish data. JRC
Working Papers on Territorial Modelling and Analysis 2019-05. Joint Research
Centre (Seville site). https://ideas.repec.org/p/ipt/termod/201905.html
[10]
Robin Burke. 2002. Hybrid Recommender Systems: Survey and Experiments. User
Modeling and User-Adapted Interaction 12, 4 (2002), 331–370. https://doi.org/10.
1023/A:1021240730564
[11]
Robin Burke. 2017. Multisided Fairness for Recommendation. arXiv:1707.00093
[cs] (July 2017). http://arxiv.org/abs/1707.00093 arXiv: 1707.00093.
[12]
R. Burke, Himan Abdollahpouri, B. Mobasher, and Trinadh Gupta. 2016. Towards
Multi-Stakeholder Utility Evaluation of Recommender Systems. In UMAP.
[13]
Laizhong Cui, Peng Ou, Xianghua Fu, Zhenkun Wen, and Nan Lu. 2017. A novel
multi-objective evolutionary algorithm for recommendation systems. J. Parallel
and Distrib. Comput. 103 (May 2017), 53–63. https://doi.org/10.1016/j.jpdc.2016.
10.014
[14]
Camélia Dadouchi and Bruno Agard. 2021. Recommender systems as an agility
enabler in supply chain management. Journal of Intelligent Manufacturing 32, 5
(June 2021), 1229–1248. https://doi.org/10.1007/s10845-020- 01619-5
[15]
Aparna Das, Claire Mathieu, and Daniel Ricketts. 2009. Maximizing prot using
recommender systems. arXiv:0908.3633 [cs] (Aug. 2009). http://arxiv.org/abs/0908.
3633 arXiv: 0908.3633.
[16]
David S. Evans and Richard Schmalensee. 2016. The New Economics of Multi-
Sided Platforms: A Guide to the Vocabulary. SSRN Electronic Journal (2016).
https://doi.org/10.2139/ssrn.2793021
[17]
Golnoosh Farnadi, Pigi Kouki, Spencer K. Thompson, Sriram Srinivasan, and Lise
Getoor. 2018. A Fairness-aware Hybrid Recommender System. arXiv:1809.09030
[cs, stat] (Sept. 2018). http://arxiv.org/abs/1809.09030 arXiv: 1809.09030.
[18]
Jesus Felipe, Utsav Kumar, Arnelyn Abdon, and Marife Bacate. 2012. Product
complexity and economic development. Structural Change and Economic Dynamics
23, 1 (March 2012), 36–68. https://doi.org/10.1016/j.strueco.2011.08.003
[19]
Luca Fraccascia, Ilaria Giannoccaro, and Vito Albino. 2018. Green product develop-
ment: What does the country product space imply? Journal of Cleaner Production
170 (Jan. 2018), 1076–1088. https://doi.org/10.1016/j.jclepro.2017.09.190
[20]
Diana Genkova. 2021. Modeling of the Human Needs: an Economic Interpretation
of Maslow’s Theory of Motivation. WSEAS TRANSACTIONS ON BUSINESS AND
ECONOMICS 18 (Jan. 2021), 253–264. https://doi.org/10.37394/23207.2021.18.26
[21]
Sally J. Goerner, Bernard Lietaer, and Robert E. Ulanowicz. 2009. Quantifying eco-
nomic sustainability: Implications for free-enterprise theory, policy and practice.
Ecological Economics 69, 1 (Nov. 2009), 76–81. https://doi.org/10.1016/j.ecolecon.
2009.07.018
[22]
Andres Gomez-Lievano. 2018. Methods and Concepts in Economic Complexity.
arXiv:1809.10781 [nlin, physics:physics, q-n] (Oct. 2018). http://arxiv.org/abs/
1809.10781 arXiv: 1809.10781.
[23]
Robert Hamwey, Henrique Pacini, and Lucas Assunção. 2013. Mapping Green
Product Spaces of Nations. The Journal of Environment & Development 22, 2 (June
2013), 155–168. https://doi.org/10.1177/1070496513482837
[24]
Ricardo Hausmann, César A. Hidalgo, Sebastian Bustos, Michele Coscia, Alexander
Simoes, and Muhammed A. Yildirim. 2013. The Atlas of Economic Complexity:
Mapping Paths to Prosperity (2nd ed.). MIT Press, Cambridge. https://mitpress.
mit.edu/index.php?q=books/atlas-economic- complexity Backup Publisher: MIT
Press.
[25]
Ricardo Hausmann and Bailey Klinger. 2006. Structural Transformation and
Patterns of Comparative Advantage in the Product Space. (2006).
[26]
C. A. Hidalgo and R. Hausmann. 2009. The building blocks of economic complexity.
Proceedings of the National Academy of Sciences 106, 26 (June 2009), 10570–10575.
https://doi.org/10.1073/pnas.0900943106
[27]
C. A. Hidalgo, B. Klinger, A.-L. Barabasi, and R. Hausmann. 2007. The Product
Space Conditions the Development of Nations. Science 317, 5837 (July 2007),
482–487. https://doi.org/10.1126/science.1144581
[28]
M. Huberty and G. Zachmann. 2011. Green exports and the global product space:
Prospects for EU industrial policy. Bruegel Working Paper 2011/07, May 2011.
[29]
Shiva Jalali, Sayyed Alireza Hashemi Golpayegani, and Hoda Ghavamipoor. 2014.
Designing a model of decision making in layers of supply, manufacturing, and
distribution of the supply chain: A recommender-based system. In 8th International
Conference on e-Commerce in Developing Countries: With Focus on e-Trust. IEEE,
Mashhad, Iran, 1–9. https://doi.org/10.1109/ECDC.2014.6836758
, Vol. 1, No. 1, Article . Publication date: September 2021.

Multiobjective recommendation for sustainable production systems •7
[30]
Dietmar Jannach and Gediminas Adomavicius. 2017. Price and Prot Awareness
in Recommender Systems. arXiv:1707.08029 [cs] (July 2017). http://arxiv.org/abs/
1707.08029 arXiv: 1707.08029.
[31]
Chen Karako and Putra Manggala. 2018. Using Image Fairness Representations
in Diversity-Based Re-ranking for Recommendations. arXiv:1809.03577 [cs] (Sept.
2018). http://arxiv.org/abs/1809.03577 arXiv: 1809.03577.
[32]
Ali Kharrazi. 2019. Resilience. In Encyclopedia of Ecology. Elsevier, 414–418.
https://doi.org/10.1016/B978-0- 12-409548- 9.10751-1
[33]
Ali Kharrazi, Brian Fath, and Harald Katzmair. 2016. Advancing Empirical Ap-
proaches to the Concept of Resilience: A Critical Examination of Panarchy, Eco-
logical Information, and Statistical Evidence. Sustainability 8, 9 (Sept. 2016), 935.
https://doi.org/10.3390/su8090935
[34]
Ali Kharrazi, Elena Rovenskaya, and Brian D. Fath. 2017. Network structure
impacts global commodity trade growth and resilience. PLOS ONE 12, 2 (Feb.
2017), e0171184. https://doi.org/10.1371/journal.pone.0171184
[35]
Qiuzhen Lin, Xiaozhou Wang, Bishan Hu, Lijia Ma, Fei Chen, Jianqiang Li, and
Carlos A. Coello Coello. 2018. Multiobjective Personalized Recommendation
Algorithm Using Extreme Point Guided Evolutionary Computation. Complexity
2018 (Nov. 2018), 1–18. https://doi.org/10.1155/2018/1716352
[36]
Weiwen Liu and Robin Burke. 2018. Personalizing Fairness-aware Re-ranking.
arXiv:1809.02921 [cs] (Sept. 2018). http://arxiv.org/abs/1809.02921 arXiv:
1809.02921.
[37]
A. H. Maslow. 1943. A theory of human motivation. Psychological Review 50, 4
(July 1943), 370–396. https://doi.org/10.1037/h0054346
[38]
Penny Mealy, J. Doyne Farmer, and Alexander Teytelboym. 2019. Interpreting
economic complexity. Science Advances 5, 1 (Jan. 2019), eaau1705. https://doi.
org/10.1126/sciadv.aau1705
[39]
Penny Mealy and Alexander Teytelboym. 2020. Economic complexity and the
green economy. Research Policy (April 2020), 103948. https://doi.org/10.1016/j.
respol.2020.103948
[40]
Rishabh Mehrotra, James McInerney, Hugues Bouchard, Mounia Lalmas, and
Fernando Diaz. 2018. Towards a Fair Marketplace: Counterfactual Evaluation
of the trade-o between Relevance, Fairness & Satisfaction in Recommendation
Systems. In Proceedings of the 27th ACM International Conference on Information
and Knowledge Management. ACM, Torino Italy, 2243–2251. https://doi.org/10.
1145/3269206.3272027
[41]
Nikola Milojkovic, Diego Antognini, Giancarlo Bergamin, Boi Faltings, and
Claudiu Musat. 2020. Multi-Gradient Descent for Multi-Objective Recommender
Systems. arXiv:2001.00846 [cs, stat] (April 2020). http://arxiv.org/abs/2001.00846
[42]
M. Otero, A. Pastor, J.M. Portela, J.L. Viguera, and M. Huert. 2011. Methods of
Analysis for a Sustainable Production System. In Climate Change - Research and
Technology for Adaptation and Mitigation, Juan A. Blanco (Ed.). InTech. https:
//doi.org/10.5772/24237
[43]
Arnault Pachot, Adélaïde Albouy-Kissi, Benjamin Albouy-Kissi, and Frédéric
Chausse. 2021. Production2Vec: a hybrid recommender system combining seman-
tic and product complexity approach to improve industrial resiliency. In 2021 2nd
International Conference on Articial Intelligence and Information Systems (ICAIIS
’21). Chongqing, China. https://doi.org/10.1145/3469213.3469218
[44]
François Perruchas, Davide Consoli, and Nicolò Barbieri. 2020. Specialisation,
diversication and the ladder of green technology development. Research Policy
49, 3 (April 2020), 103922. https://doi.org/10.1016/j.respol.2020.103922
[45]
Christian Reynolds, Manju Agrawal, Ivan Lee, Chen Zhan, Jiuyong Li, Phillip
Taylor, Tim Mares, Julian Morison, Nicholas Angelakis, and Göran Roos. 2018. A
sub-national economic complexity analysis of Australia’s states and territories.
Regional Studies 52, 5 (May 2018), 715–726. https://doi.org/10.1080/00343404.
2017.1283012
[46]
Marco Tulio Ribeiro, Nivio Ziviani, Edleno Silva De Moura, Itamar Hata, Anisio
Lacerda, and Adriano Veloso. 2015. Multiobjective Pareto-Ecient Approaches for
Recommender Systems. ACM Transactions on Intelligent Systems and Technology
5, 4 (Jan. 2015), 1–20. https://doi.org/10.1145/2629350
[47]
Jean-Charles Rochet and Jean Tirole. 2003. Platform Competition in Two-Sided
Markets. Journal of the European Economic Association 1, 4 (June 2003), 990–1029.
https://doi.org/10.1162/154247603322493212
[48]
Barry Smyth and Paul McClave. 2001. Similarity vs. Diversity. In Case-Based
Reasoning Research and Development, G. Goos, J. Hartmanis, J. van Leeuwen,
David W. Aha, and Ian Watson (Eds.). Vol. 2080. Springer Berlin Heidelberg,
Berlin, Heidelberg, 347–361. https://doi.org/10.1007/3- 540-44593- 5_25 Series
Title: Lecture Notes in Computer Science.
[49]
Robert E. Ulanowicz. 2003. Some steps toward a central theory of ecosystem
dynamics. Computational Biology and Chemistry 27, 6 (Dec. 2003), 523–530.
https://doi.org/10.1016/S1476-9271(03)00050- 1
[50]
Robert E. Ulanowicz, Sally J. Goerner, Bernard Lietaer, and Rocio Gomez. 2009.
Quantifying sustainability: Resilience, eciency and the return of information
theory. Ecological Complexity 6, 1 (March 2009), 27–36. https://doi.org/10.1016/j.
ecocom.2008.10.005
[51]
Robert E. ULANOWlCZ and Jerey S. Norden. 1990. Symmetrical overhead in
ow networks. International Journal of Systems Science 21, 2 (Feb. 1990), 429–437.
https://doi.org/10.1080/00207729008910372
[52]
The Growth Lab At Harvard University. 2019. International Trade Data (HS, 92).
https://doi.org/10.7910/DVN/T4CHWJ Type: dataset.
[53]
The Growth Lab At Harvard University. 2019. Product Space Networks. https:
//doi.org/10.7910/DVN/FCDZBN Type: dataset.
[54]
Guido van Capelleveen, Jesse van Wieren, Chintan Amrit, Devrim Murat Yazan,
and Henk Zijm. 2021. Exploring recommendations for circular supply chain
management through interactive visualisation. Decision Support Systems 140 (Jan.
2021), 113431. https://doi.org/10.1016/j.dss.2020.113431
[55]
Shanfeng Wang, Maoguo Gong, Haoliang Li, and Junwei Yang. 2016. Multi-
objective optimization for long tail recommendation. Knowledge-Based Systems
104 (July 2016), 145–155. https://doi.org/10.1016/j.knosys.2016.04.018
[56]
Sirui Yao and Bert Huang. 2017. New Fairness Metrics for Recommendation that
Embrace Dierences. arXiv:1706.09838 [cs] (Dec. 2017). http://arxiv.org/abs/1706.
09838 arXiv: 1706.09838.
[57]
Yi Zuo, Maoguo Gong, Jiulin Zeng, Lijia Ma, and Licheng Jiao. 2015. Personalized
Recommendation Based on Evolutionary Multi-Objective Optimization [Research
Frontier]. IEEE Computational Intelligence Magazine 10, 1 (2015), 52–62. https:
//doi.org/10.1109/MCI.2014.2369894
, Vol. 1, No. 1, Article . Publication date: September 2021.