This is a peer-reviewed, accepted author manuscript of the following paper : Srinual, N., Mehnen, J., & Boonraksa, N.

(Accepted/In press). A review of

sustainable supplier selection and evaluation using topic modelling. Paper presented at The 8th HCU National and International Conference,
Samutprakarn, Thailand.

A Review of Sustainable Supplier Selection and Evaluation Using

Topic Modelling
Napat Srinual1, Jörn Mehnen2, and Nantawan Boonraksa3

1,2Department of Design, Manufacturing and Engineering Management, Faculty of Engineering,

University of Strathclyde, Glasgow,UK
3Department of Logistics and Supply Chain Management, Faculty of Business Administration,

Huachiew Chalermprakiet University, Samutprakarn, Thailand

Email: napat.srinual@strath.ac.uk Tel. +44 759-786-8890

Abstract
The study provides a review of sustainable supplier selection (SSS), which has
proceed to emerge and expand along with greater speed over the last decades. This
research aims to identify and explore the use of machine learning and text mining methods
to automatically identify relevant core topics in sustainable supplier selection. The
traditional approach by personal judgments with predetermined categories, cannot
adequately take latent topics from large volumes of research data. Therefore, this study
selects the topic modelling approach, which automatically identify topics that extend a
large and unstructured collection of documents to uncover research topics in sustainable
supplier selection research. The papers for the literature review were collected from
SCOPUS database. The model with 20 topics was selected through the Latent Dirichlet
Allocation (LDA) model from selected articles published from 2010 to 2020 in associated
with various journals, and the top 5 most popular topics in sustainable supplier selection
research are reviewed in a nutshell. We then explore topic trends by considering the
transformation over different topics of sustainable supplier selection research over the last
few decades. For each of the top 10 journals, the areas of subspecialty and the effects of
editor changes on the topic portfolios are also investigated. The findings of this study are
expected to provide implications for researchers, journal editors, and policy makers in the
field of sustainable supplier selection.

Keywords: Sustainability; Supplier selection and evaluation; Latent Dirichlet Allocation;

Topic modelling; Unstructured documents
1. Introduction
Corporate sustainability management is not limited to a single company. The
importance of the influence of suppliers and the supply chain on the sustainability of
products and processes is now undisputed. (Mani, Gunasekaran, & Delgado, 2018).
Considering supplier practices and characteristics, determining the best sustainability
match for those suppliers, and then selecting those suppliers are both important and
accepted by practitioners and researchers to achieve truly sustainable outcomes
(Shahryari et al. 2016).
In addition, both the literature review and the review papers on the problem of
sustainable supplier selection (SSSP) have grown at a rapid pace and have become an
academic discipline in their own right (Bai & Sarkis 2014). The SSSP literature has
expanded significantly with more than a dozen academic journals devoted to the field.
The substantial growth of the research has spurred bibliometric studies on itself because
it is common for scholars to turn their attention to the literature itself once an academic
discipline reaches a certain level of maturity (Chai, Liu, & Ngai, 2013). For example, 143
papers were selected for a literature review on supplier selection in sustainability.
However, this is a time-consuming process with limited processing power, resulting in a
small number of papers analysed. Researchers, especially early career researchers, often
need to find, organise, and understand new and unexplored areas of research. Since a
literature review often involves a large number of papers at the initial stage, the option
for a researcher is either to limit the number of papers to be reviewed a priori or to review
papers using other methods. It is interesting to note that there is another literature review
related to sustainable supplier selection, e.g. Genovese et al. (2013) reviewed quantitative
models to support green supplier selection. Moreover, these studies use the conventional
discrete task approach, where each article is classified into a single category of
predetermined topics based on subjective judgement. However, such discrete tasks cannot
effectively hold latent themes from large scientific data. For one, manual classification
based on reading abstracts or author keywords always carries the risk of classification
error. Moreover, it costs too much time, especially when the number of articles to be
assigned is very large. Second, the given categories are by no means exhaustive; relatively
new and emerging topics are likely to be ignored. Convergence topics are also difficult to
handle. Third, an article usually contains two or more topics. So far, handling large
collections of articles has been structured into topics or categories by using coding sheets
(Antons, Kleer, & Salge, 2016), dictionaries, or supervised learning methods. After
considering some weaknesses in the existing literature reviews, we attempt to build a
comprehensive literature review on supplier selection for sustainability from the database
SCOPUS. This work applies topic modelling (Blei, Ng, & Jordan, 2003) called Latent
Dirichlet Allocation (LDA). The period of data collection is 2010 to 2020.

2. Objectives
To review on topic trends, differing thematic fields and emerging research gap
areas in the topic in sustainable supplier selection and evaluation by applying LDA topic
modelling technique.

3. Materials and methods

This section clarifies the methods and procedure for distinguishing themes or
topics in sustainable supplier selection research through LDA topic modelling. The
procedure involves three steps: (1) target journal selection and data collection, (2) text
pre-processing, (3) topic modelling, as shown in Figure 1

Figure 1 Procedure for identifying topics and applying topic modelling

3.1 Journal pre-selection in SCOPUS

The first step is to select relevant and leading journals for the article collection of
the research on sustainable supplier selection (SSS) problems. Figure 2 shows that the
number of articles has increased sharply from 2010 to 2020. This is the significant
evidence based on the frequency of distribution by year of publication in sustainable
supplier selection.

1044

325
182

1990 1995 2000 2005 2010 2015 2020

No. of publication

Figure 2 The numbers of published paper in sustainable from 2010 to 2020

The selection of leading journals in the field of sustainable supplier selection

(SSS) is influenced by many aspects of SSS, but recent studies looking at SSS literature
provide evidence on journal selection. Some recent studies confirmed that this journal can
be considered as one of the leading journals due to its history and relatively high impact
compared to other journals.
Table 1 lists the selected 10 journals and the corresponding number of articles
included in this study.

Table 1 Lists of Sustainable Supplier Selection (SSS) Journals

Journal and conference Year started Impact Factor Number
Journal Of Cleaner Production 1993 7.246 325
Sustainability Switzerland 2009 2.576 307
Iop Conference Series Earth And 2008 0.45 65
Environmental Science
Energy 1976 6.082 58
Science Of The Total Environment 1972 6.551 55
Energies 2008 2.702 48
Advances In Intelligent Systems And 2012 0.57 42
Computing
Journal Of Environmental Management 1999 4.175 41
Applied Energy 1999 8.848 37
Iop Conference Series Materials Science 2009 0.53 37
And Engineering
 For each journal, the in-depth process is carried out to gather the raw data for the
targeted articles. First, documents that are original articles and reviews are examined in
the database SCOPUS using different keywords for each journal starting from 2010 to
2020 as shown in Table 2.

Table 2 Identifying keywords in source database

Keywords
Sustainable Supplier selection and evaluation
supplier, selecti*, evaluate*
sustainabl*
vendor, selecti*, evaluat*

Then, five meta-variables, namely title, year, DOI, abstract and keywords are
selectively downloaded for the selected articles. In total, we collected 1,015 articles
published in the top ten journals from 2010 to 2020. Table 3 shows the first three rows of
my selected article data below.

Table 3 Sample of first three rows from selected articles

Title Year DOI Abstract Keywords
An Approach to Develop 2020 10.1088/1757- In construction industry, -
a Sustainable Preference 899X/ selecting a right material
Index for Self 998/1/012058 from the various alternatives
Compacting Concrete is a critical task in
contributing to promote
sustainability. For
evaluating performance of
the material different
conflicting attributes are
considered which is very …
Decision-making 2020 10.1088/1755- The article describes the Analytic hierarchy
information support 1315/ methodology of decision- process; Application
methodology for 613/1/012073 making information support programs;
protective measures for protective measures Biotechnology; Crops;
implementation in the implementation in the agro- Engineering education;
AIC industrial complex (AIC). Environmental
The aim is to determine the management;
rational area of treated land Environmental
plots for protective measures technology …
implementation as …
An integrated model for 2020 10.1016/ In today's competitive BWM; PROMETHEE;
selecting suppliers on the j.jclepro. business environment, Supplier selection;
basis of sustainability 2020.123261 corporations attempt to Sustainable innovation;
innovation achieve sustainability Sustainable supply
through innovation. chain management …
Innovation is considered by
researchers and scholars to
be a key driver for achieving
sustainability…

3.2 Text pre-processing

Since we load the selected papers into Google Colab (python programming) , we
applied ‘title , ‘abstract’ and ‘keywords’ column of each paper to handle them as a single
document to form a new column called ‘Article text’. Python code and ‘Article text’
column can be seen in Figure 3 and Figure 4 below.
 Figure 3 Load and merge a raw text data in Google Colab

Figure 4 A sample of “Article text” data

In the text pre-processing method, a course of cleaning processes (tokenization,

lowercase, etc.) to exceptional results of documents is then presented in Figure 5.
 Figure 5 Overview of cleaning process in text pre-processing

The cleaning process is usually a repetitive one, as it can be challenging to locate

the entire misinterpreted and meaningless words. The corpora of several papers consist
of different words, which means that an accurate cleaning process cannot be guaranteed
when a new review is performed. Therefore, developing features and applying text mining
packages for cleaning raw text can contribute to a better cleaning process. Figure 6 shows
the process of text cleaning in an example article.

Figure 6 Text-cleaning process sample

 After text preprocessing, the cleaned and preprocessed text data was inserted into
a new column in the data frame labelled "Clean Article Text" and then de-tokenized. The
example from the "Clean Article Text" column is shown in Table 4.

Table 4 Comparison of documents current-former cleaning process

No. Article text (former) Clean article text (current)
1 An Approach to Develop a Sustainable Preference approach develop sustainable preference index
Index for Self Compacting Concrete in construction self compact concrete in construction industry
industry, selecting a right material from the various select right material various alternative critical
alternatives is a critical task in contributing to task contribute promote sustainability …
promote sustainability…
2 The aim is to determine the rational area of treated determine rational area treat land plot protective
land plots for protective measures implementation as measure implementation component cultivate
components of cultivated crop area-based industrial crop area based industrial technology …
technology …

The final step of preprocessing is to transform the documents into the document
term matrix representation to generate the input for LDA. As input to LDA, the
preprocessed corpus was aligned to a document term matrix (DTM). A document term
matrix is a mathematical matrix that describes the frequency of terms that occur in a
collection of documents. Figure 7 shows how the usage of some user specified words
changes over time. It also represents the number of each term per document in Table 5.
To do this, the Gensim package can generate a unique ID for each word in the document.

Figure 7 Trend of % frequency of sample word from 2010 to 2020

Table 5 lists the 20 words with the highest frequencies, as well as words taken
from the Clean Article Text column.
 Table 5 Top 20 no. of occurrences from ‘Clean article text’ in 1,150 papers
Rank Words Frequency Rank Words Frequency
1 sustainability 1704 11 fuzzy 524
2 method 1599 12 policy 444
3 analysis 992 13 stakeholder 433
4 energy 937 14 cost 431
5 criterion 751 15 optimization 424
6 environmental 718 16 construction 398
7 green 654 17 ranking 352
8 decision making 595 18 manufacturing 237
9 design 586 19 ahp 227
10 evaluation 557 20 water 209

Then, we create a wordcloud to see the most common words used in papers from
2010 to 2020, as shown in Figure 8.

Figure 8 Summary a frequent word in Wordcloud

3.3 Topic modelling approach

Topic models have several advantages, both from an algorithmic and a practical
point of view. First, they have principled mathematical foundations that help us identify
the mechanism of document structure. Second, they do not require prior labeling of
documents, so documents can be viewed without the help of expert knowledge (Blei et
al. 2003). Finally, they can shape and organize documents without any doubt (Griffiths
and Steyvers 2004). Due to these advantages, topic models have gained new importance
and have been successfully applied to a wide range of text mining tasks. LDA, proposed
by Blei (2012), is the most widely used topic modeling algorithm and has several practical
advantages over other topic modeling algorithms. It can classify the combination of
existing topics in new documents without updating the current model. The key idea of
LDA is that documents are specified as random mixtures over latent topics, each of which
is defined by a distribution over words.
Therefore, once the papers have been cleaned, the LDA method can be operate.
To build the LDA model, these settings consist of three key parameters alpha (α), beta
(β) and the number of topics (K). Alpha (α) estimated the topic distribution over
documents, Beta (β) determines the specificity of topics. To build LDA model, we
employed alpha = 50/K, beta = 0.1 and the number of topics (K) = 20.
4. Results
The LDA results from Google Colab in Table 6 show 20 topics, each having 10
keywords, in 1,015 papers.

Table 6 20 topics in sustainable supplier selection (1,015 papers)

No. of
Topic # 00 Topic # 01 Topic # 02 … Topic # 17 Topic # 18 Topic # 19
Keywords
0 model supplier energy … china food concrete
supplier renewable … power food waste construction
1 method
selection energy
… power local aggregate
2 analysis industry electricity
generation
developmen … performance production material
3 suppliers biomass
t evaluation
. . . . . . . .
. . . . . . . .
. . . . . . . .
approach procurement energy life cycle evaluate chain
8 …
efficiency
criterion criterion electricity materials operation chains
9 …
generation

Then, the LDA results from 1,015 papers are viewed in Figure 9. The closer the
distance among the circles, the more related topics with words among the circles.

Figure 9 20 topics of papers by LDAvis results

Figure 9 shows the two major groups with a small distance between the circles.
The first group consists of topic1, 4, 6, 8, 11, 12 and 16. There is a relationship in decision
making analysis with linguistics, fuzzy logic, which has increased in the context of
sustainable supplier selection techniques. In addition, China is still leading the research
on sustainable supplier selection. The second group is topic 2, 3, 19 and 14, which focuses
on the industry in terms of environmental impact, including renewable energy, biomass,
electricity and food waste from production to select sustainable suppliers. However, there
is a positive trend in the automotive industry to apply supplier selection for sustainability.
In addition to the two topic areas, it is interesting to note that topic 18 deals with risk
assessment, which stays away from the other topic areas. This does not mean that research
with such topics is not necessary or weak. Rather, it may be an emerging research topic.
In particular, the research topic of findings could be triggered by them.

4.1 Clusters by “Clean article text”

Table 7 shows five clusters consisting of the following keywords among 1,015
articles associated with each cluster, as shown in Figure 10.

Table 7 Five clusters from LDAvis results

Cluster Term for each topic
1 decision making, multicriteria analysis, fuzzy logic, linguistic, criterion, economic, China, ahp,
anp
2 design, construction, building, material, eco-design, concrete, production, manufacturing
3 urban, recycling, waste management, social impact, automotive, delivery
4 green, environmental, renewable energy, energy consumption, optimization, electricity, food
waste
5 soil, chemical, land use , agriculture, water management

Figure 10 Cluster labelling from LDAvis

4.2 Publication papers on economy country conditions

Distribution of published papers based on country conditions is an interesting
topic in our review paper. In this section, based on developed and developing countries
we showed that, which economy region used decision making approaches in sustainable
supplier selection in Figure 11. Still, Thailand, there is lack of research in this field.
 Figure 11 Distribution of papers based on economy country conditions

4.3 Publication by country

We also analyzed the distribution of reviewed papers by country of origin. Table
8 shows the countries in the field of sustainable supplier selection and evaluation. As can
be seen, more than 70% of the publications (208 publications) are from four countries
(China, Iran, India and United States), where China is the leading country in terms of
number of publications with 73 publications, followed by Iran with 54 publications. On
the other hand, there are only a few publications in Thailand that deal with sustainability
in supplier selection.

Table 8 Distribution of papers based on country (Top 10)

Country Number Percentage (%)
China 73 22.26
Iran 54 16.46
India 45 13.72
United States 36 10.98
Taiwan 29 8.84
United Kingdom 23 7.01
Malaysia 21 6.40
Germany 20 6.10
Australia 14 4.27
Turkey 13 3.96

4.4 Publication by Multicriteria decision making approaches (MCDM)

In this study, we summarize two categories used for sustainable supplier
evaluation and selection, namely the single approach and the hybrid approach.

Table 9 the classification of MCDM techniques

Single Percentages Hybrid Percentages
approach approach
AHP 24.29 AHP, TOPSIS 36.04
ANP 21.43 AHP, ANP 12.61
TOPSIS 34.29 ANP, ELECTRE 10.81
VIKOR 4.29 AHP, PROMETHEE 10.81
ELECTRE 4.29 Fuzzy AHP, Fuzzy TOPSIS 22.52
PROMETHEE 7.14 Fuzzy ANP, Fuzzy TOPSIS 3.60
BWM 2.86 Fuzzy ANP, DEMATEL 1.80
MAUT 1.43 Fuzzy DEMATEL, Fuzzy VIKOR 1.80
 Table 9 shows that about 80% of the single or individual approaches refer to AHP,
ANP and TOPIS. As a hybrid or combined approach, AHP, ANP and TOPSIS are
introduced to integrate between these techniques. Nowadays, most authors focus on
modelling approaches that support combined multi-criteria decision making rather than
single approaches. However, in sustainable supplier evaluation and selection, there is a
gap between using single and hybrid approaches to solve this problem. Therefore,
integrating fuzzy logic to create a novel approach in MCDM techniques is still a space
for research and overcome the shortcomings of traditional methods.

5. Conclusion
This study identified 20 sustainable supplier selection research topics and
examined their changes and relationships by applying the LDA theme modelling
approach to 1,015 articles published in the SCOPUS database over the past decade. The
findings on the research trends in the articles during the period of 2010-2020 are
summarized in five key trend themes. First, the majority of research topics in sustainable
supplier selection deal with a multi-criteria decision-making approach. A single approach
has a negative trend, while a hybrid MCDM approach has a positive trend over the past
10 years. Second, although the construction industry mostly applies sustainable supplier
selection, the automotive industry has increased the trend to focus on sustainable supplier
selection. Third, carbon emissions and renewable energy still play an important role, not
only in economic and environmental terms, but also in terms of social impact for the
welfare of people in the future generation. The fourth theme shows that circular economy
trends have a significant impact on sustainability, such as recycling and waste
management. Finally, the impact of land use in agriculture has also increased the trends
in sustainability. Nevertheless, this study also has some limitations that could serve as
areas for future research. The primary limitation is that we show the themes for the entire
10-year period. We cannot discover the evolution of the themes using such a cumulative
approach. Next, we applied the basic LDA model to discover topics and their popularity
trends over time. While there are LDA extensions, these have been evolved. Finally, we
put only 20 topics through the majority of topic modelling approach. For further research,
expanding the scope of research in this way would support explicit consideration for
sustainability transformation.

6. References
Antons, D., Kleer, R., & Salge, T. O. (2016). Mapping the topic landscape of JPIM, 1984–
2013. In search of hidden structures and development trajectories. Journal of
Product Innovation Management, 33(6), 726–749.
Bai, C., & Sarkis, J. (2014). Determining and Applying Sustainable Supplier Key
Performance Indicators. Supply Chain Management: An International Journal, 19
(3), 275–291.
Blei, D.M., Ng, A.Y., & Jordan, M.I. (2003). Latent dirichlet allocation. Journal of
Machine Learning Research, 3, 993–1022.
Blei, D.M. (2012). Probabilistic topic models. Communications of the ACM, 55(4), 77–
84.
Chai, J., Liu N.K., & Ngai, W. T. (2013). Application of Decision-making Techniques in
Supplier Selection: A Systematic Review of Literature. Expert Systems with
Applications, 40(10), 3872–3885.
Genovese, A., Lenny Koh, S., Bruno, G., & Esposito, E. (2013). Greener supplier
selection: state of the art and some empirical evidence. International Journal of
Production Research, 51(10), 2868-2886.
Griffiths, T.L., & Steyvers, M. (2004). Finding scientific topics. Proceedings of the
National Academy of Sciences, 101, 5228–5235.
Mani, V., Gunasekaran, A., & Delgado, C. (2018). Enhancing supply chain performance
through supplier social sustainability: an emerging economy perspective.
International Journal of Production Economics, 195, 259-272.
Shahryari Nia, A., Olfat, L., Esmaeili, A., Rostamzadeh, R., & Antuchevičienė, J. (2016).
Using fuzzy Choquet Integral operator for supplier selection with environmental
considerations. Journal of Business Economics and Management, 17(4), 503-526
Tracey, M., & Leng Tan, C. (2001). Empirical analysis of supplier selection and
involvement, customer satisfaction, and firm performance. Supply Chain
Management: An International Journal, 6(4), 174-188.
Zhou, X., & Xu, Z. (2018). An Integrated Sustainable Supplier Selection Approach Based
on Hybrid Information Aggregation. Sustainability, 10(7), 2543.