Advances in AI-Based Garment Returns Prediction and Processing:
A Conceptual Approach for an AI-Based Recommender System
Soeren Gry
a
, Marie Niederlaender
b
, Aena Nuzhat Lodi, Marcel Mutz
c
and Dirk Werth
d
August-Wilhelm Scheer Institut, Uni Campus D 51, Saarbr
¨
ucken, Germany
Keywords:
Recommender System, Sustainable Return Management, Logistics, Sustainable Supply Chain, E-Commerce,
Fashion, Apparel, Artificial Intelligence, Machine Learning.
Abstract:
The ever-increasing volume of returned garments not only represents a huge increase in costs for retailers or
manufacturers and an inventory risk that is difficult to calculate, but also has a high environmental impact
due to the destruction of the garments and the necessary logistics processes. Most of the existing solutions to
these problems aim to eliminate returns altogether. However, many returns cannot be avoided, e.g. orders for a
selection of products, repairs or quality-related returns. For this reason, this study explores the potential of AI
to predict returns and make recommendations on how best to plan the reverse logistics network, resulting in
environmental and economic benefits. To this end, an extensive literature review was conducted to capture the
current state of research. Based on this, a conceptual approach for the development of an AI-based recommender
system for the best possible handling of returns will be derived.
1 INTRODUCTION
During the corona pandemic, the share of global e-
commerce rose from 15% of total retail sales in 2019
to 21% in 2021. For 2022, a share of 22% of total re-
tail sales is expected. As a result, e-commerce growth
has slowed but remains high and is expected to in-
crease from $3.3 trillion in 2022 to $5.4 trillion in
2026 (Morgan Stanley, 2022). The volume of trans-
port and product returns increases accordingly. While
e-commerce has significant advantages for consumers
over bricks-and-mortar stores, and not just during a
global pandemic, there are also some disadvantages. E-
commerce does not allow consumers to see, feel or try
products before they buy them. Returns are therefore
inevitable (Asdecker and Karl, 2018). In a European
comparison, Germany has the highest number of par-
cel returns. Around 24% of parcels in Germany are
returned by consumers. In 2021, around 530 million
returns involving 1.3 billion items were made in Ger-
many. 91% of returned items were apparel and shoes.
As well as the economic harm caused by the difficulty
of achieving positive margins on e-commerce, the envi-
a
https://orcid.org/0000-0002-4441-0517
b
https://orcid.org/0009-0008-1935-821X
c
https://orcid.org/0000-0002-5918-6407
d
https://orcid.org/0000-0003-2115-6955
ronmental consequences of this ordering behaviour are
also significant. An estimated 795,000 tonnes of CO2
are caused by returns in Germany alone, the equiva-
lent of 5.3 billion car kilometres (Forschungsgruppe
Retourenmanagement, 2022). Returns are one reason
why the fashion and apparel industry is responsible for
5% of global emissions, more than international air
travel and cruise ships combined. This makes the fash-
ion and apparel industry one of the top three polluters
and puts it at the centre of policy efforts (Vogue/BCG,
2021). The consumer behaviour described in relation
to returns is not only observed in Germany, where
up to 60% of apparel and shoes are returned, but is
also almost congruent at European level (Forschungs-
gruppe Retourenmanagement, 2022). Manufacturers
and retailers take preventive and reactive approaches to
dealing with returns (Deges, 2021). This study looks
at current reactive approaches to reducing the envi-
ronmental and economic costs of returns. In order to
handle returns in the best possible way and to recycle
them in the sense of second life planning, manufactur-
ers and retailers depend on accurate forecasts of the
volume of returns. In this context, Artificial Intelli-
gence (AI), and in particular Machine Learning (ML)
as part of AI research, can be used to recognise pat-
terns and regularities in huge data sets (Lickert et al.,
2021; Schwaiger and Steinwendner, 2019). This re-
search paper examines current scientific approaches to
Gry, S., Niederlaender, M., Lodi, A., Mutz, M. and Werth, D.
Advances in AI-Based Garment Returns Prediction and Processing: A Conceptual Approach for an AI-Based Recommender System.
DOI: 10.5220/0012010500003552
In Proceedings of the 20th International Conference on Smart Business Technologies (ICSBT 2023), pages 15-25
ISBN: 978-989-758-667-5; ISSN: 2184-772X
Copyright
c
 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)
15

1) predicting returns and 2) providing them with the
best possible processing. The focus of this paper is on
AI-based approaches and the identification of current
research gaps. At the same time, priority is given to
literature focused on the fashion and apparel sector
due to the high environmental impact of this industry.
The structure of the research paper is as follows. First,
AI-based approaches are explored to predict returns,
return quantities and reasons for returns. Secondly,
the AI-based processing of returns within the reverse
logistics process is investigated, with the aim of en-
abling economically and ecologically sensible further
processing. All findings from these sections are sum-
marised in Table 1. Then the paper discusses the use
of an AI-based recommender system that can be part
of reverse logistics planning. The final section sum-
marises the findings, including identified research gaps,
and provides an outlook.
2 METHODOLOGY
This paper aims to give an overview of the most recent
practices and innovations in the area of return manage-
ment with a focus on the fashion and apparel industry,
covering the topics (Figure 1):
1. Returns Forecasting and Consumer Returns Behav-
ior
2. Reverse Logistics Network Design and Optimisa-
tion (including Returns Management)
The literature review focused primarily on the five-year
period between 2018 and 2022, and was conducted be-
tween October 2022 and February 2023. For that mat-
ter Google Scholar, Wiley Online Library, ScienceDi-
rect and Springer Link were searched, including snow-
balling techniques to include relevant older sources
outside our primary search period. This paper repre-
sents an overview of the most relevant developments
in the fashion and apparel industry with a roadmap
for future research, it does not aim to provide a sys-
tematic review on return management. The following
search keywords were used for Topic 1: Returns Fore-
casting, Product Return Prediction, Consumer Return
Behavior, Prediction of Consumer Returns; and for
Topic 2: Reverse Logistics Network Design, Reverse
Logistics Network optimisation, Reverse Logistics,
Returns Management and for Context: Sustainability,
AI, Artificial Intelligence, Machine Learning, Fashion,
Apparel, Online Retail, E-Commerce. New or partic-
ularly relevant publications that relate to, or could be
translated into, the fashion and apparel industry were
selected.
Figure 1: Structure of topics along the supply chain with
top-down increasing detail level.
3 LITERATURE REVIEW
3.1 On Return Forecasting and
Observations of Return Behavior
3.1.1 Comparison of Machine Learning
Methodologies
Mathematical models have been proven to be diffi-
cult to establish in multi-parameter and multi-variable
problems (T
¨
uyl
¨
u and Ero
˘
glu, 2019). A problem such
as prediction of returns can have many possible factors,
and so in order to explore all such possibilities, many
recent studies have covered a variety of techniques
and implications regarding the use of ML methods to
forecast returns in fashion and apparel markets.
A Preliminary Selection of Algorithms: T
¨
uyl
¨
u and
Ero
˘
glu (2019) performed one such analysis and com-
parison of different ML methods to estimate a demand
forecast, which they were then able to extrapolate into
an estimate of product returns. Their experiment took
into account algorithms in four distinct categories:
lazy, rule-based, decision tree, and functional. The
dataset consisted of information on fashion and ap-
parel products, primarily women’s trousers. The high-
est levels of accuracy obtained on this large dataset
were consistently achieved by the M5P decision tree
algorithm, which combines the underlying principles
behind multiple linear regression and decision trees
for data mining (Quinlan et al., 1992). Besides this,
among the functional algorithms, linear regression and
support vector regression also proved to be a good fit
for the problem, whereas multilayer perception did
not perform as well as the others. Both of the eval-
uated rule-based algorithms, M5Rules and Decision
Table, demonstrated comparable performance as well.
M5Rules has the ability to make predictions for nom-
inal and numerical quantities, by selecting the most
effective rule generated in each cycle of tree creation
(Holmes et al., 1999).
ICSBT 2023 - 20th International Conference on Smart Business Technologies
16

Simple Data Mining Methods: In the study by As-
decker and Karl (2018), the performances of simple
and complex data analysis methods for the purpose of
predicting consumer returns were compared and con-
trasted, to determine whether or not the use of complex
methods is required to predict returns by customers.
Instead of customer information such as order and re-
turn history or shopping basket composition, they were
able to make use of shipment and returns information.
It was shown that even the simplest methods such as
binary logistic regression and linear discriminant anal-
ysis did not lag so far behind more complex methods
such as ensembles, helping retailers to identify the
variables that contribute to product returns. Return
probability was found to be positively correlated with:
the total value of the shipped goods within a package,
the number of items in a shipment, and the age of the
account that the consumer used to order the products;
and negatively correlated with the delivery time. The
authors also found that packages delivered to women
have the highest probability of being returned.
Feature Extraction for Large Sparse Datasets: In
order to create high-quality ML models that produce
accurate return forecasts, large datasets with informa-
tive features are absolutely crucial. Urbanke et al.
(2015) presented Mahalanobis feature extraction, their
newly developed method to reduce dimensionality of
large-scale sparse datasets whilst retaining useful in-
formation from the dataset. The authors carried out
their experiment on a dataset from a leading fashion
retailer in Germany that had a returns rate of 57.3%.
Using a sparse matrix format, the memory requirement
of the dataset was reduced by more than 99%. Ma-
halanobis feature extraction outperformed all of the
seven other feature extraction methods that it was com-
pared with in the study, demonstrating its utility with
regards to such datasets. Besides this, the combination
of Mahalanobis feature extraction and adaptive boost-
ing was also outperforming against logistic regression
and linear kernel support vector machine. The hybrid
method was able to predict the sales which had a very
high probability of being returned. Further work on
such an approach may prove fruitful in efficiently and
accurately processing returns data for predictions.
Time-Series Model: Return forecasting is essen-
tially a time-based problem, and so, a model based on
time-series or lagged sales may be a good fit. Shang
et al. (2020) explore this perspective, showing that
such models may indeed decrease the prediction error
by up to 18% with the right configurations. The predic-
tions were made using a predict-aggregate approach on
data from an online jewelry retailer, that formally ac-
cepted returns within 30 days after purchase, but also
did not often reject late returns. Overall, the lagged
sales regression model outperformed the time-series
based ARIMA model (Jenkins, 1970) in most of the
categories that were measured.
3.1.2 Findings Regarding Return Behaviour
Returns occur due to many reasons, including but not
limited to impulse buying due to sales or promotions,
mismatch with expectations, or intentionally, when
the customer orders a larger set of items than they
plan to retain. The following section describes the
findings in literature pertaining to return forecasting
using ML techniques, mainly in the industry of fashion
and apparel.
Demographics and Return Rates: Makkonen et al.
(2021) explored the implications of consumer demo-
graphics and preferences in payment methods on prod-
uct returns made by online consumers. The authors
collected data via a questionnaire aimed at online con-
sumers in Finland, asking the participants about their
gender, age, education and income, along with their
payment method preference, and return behaviour. Us-
ing this information, a three-part analysis of the data
was carried out. Firstly, the product return frequency
was correlated using cumulative odds ordinal logis-
tic regression. Secondly, the reasons for returning
items stated by participants were categorised via con-
tent analysis. Thirdly, specific product return reasons
were learned using binomial logistic regression. Re-
turns were found to be made more often by women,
younger people, and people who preferred to pay via
invoice. 12 participants, of whom most were in their
20s, brought up the custom of bracketing. Bracketing
refers to the purchasing of multiple items, typically
within the same order, with the intention of keeping
only a subset of them and returning all the rest. Brack-
eting is most commonly carried out with fashion items,
primarily apparel and shoes (Bimschleger et al., 2019).
When analysing the reasons for product returns, the
authors observed that over 60% of responses cited re-
turning clothes and shoes due to wrong size or bad
fitting, and also that these returns were most often
made by women. On the other hand, returns made due
to a faulty or damaged product were mostly returned
by men, and also by people who very rarely return
products (less than yearly). People who return prod-
ucts as frequently as returning monthly were much
more likely to state that they returned products due to
a mismatch with their own expectations of the items.
Advances in AI-Based Garment Returns Prediction and Processing: A Conceptual Approach for an AI-Based Recommender System
17

Consumer Behaviour Leading to Returns: Rea-
sons for consumer returns have also been closely exam-
ined by Asdecker et al. (2017), with the use of linear
and logistic regression using data from a German on-
line shop dealing mainly in women’s apparel. This
study included information on customers’ shopping
baskets, order and return history, as well as payment
method, coupons, and whether or not a free gift was
included for each order. This study finds more specific
correlations with regards to bracketing. Another such
distinct and perhaps unexpected finding is that order-
ing the same item in multiple colours in fact lowers
the return rate. The higher the average product price of
the order, the higher the likelihood of a return. Paying
for an order by invoice also increases the chance of a
return, similar to the finding by Makkonen et al. (2021)
regarding payment method preference. Returns were
also found to be associated with the use of coupons,
which may cause customers to make risky purchases
on impulse. Including a free gift in the package was
also found to reduce the probability of a return from
such an order. Lastly, the greatest impact was found
using the historical returns information pertaining to
each article and each customer. Both of these return
rates are directly proportional to the probability of the
return of an item by a customer, providing the most
information to the predictive model.
Pricing and Product Reviews: Sahoo et al. (2018)
have found that apparel products with a lesser num-
ber of reviews increase the likelihood of bracketing
by customers. When products sold online have mul-
tiple unbiased reviews from customers, their return
rate lessens. Unbiased ratings mean that there are no
incorrect ratings or, for example, conflicts of interest
on the part of the raters. Conversely, biased reviews
on a product are associated with increased return rates.
They also found that expensive items are less likely
to be returned, probably because the customer has put
more mental effort into considering their decision be-
fore making the purchase. The approach adopted by
the authors makes use of an analytical model which
they propose, and a two-stage Probit model, a type of
binary regression model (Heckman, 1979), to approx-
imate the effect of product reviews on purchases as
well as returns.
Shipping Practises: Free shipping promotions have
been shown to positively influence spending on prod-
ucts that are difficult for customers to assess, which
in turn increases the return rate (Shehu et al., 2020).
It is important to note that this does not include free
shipping policies. Shehu et al. (2020) explain that
customers view such promotions as compensation for
possible returns. The study makes use of a Type II
tobit model, in which some part of the target variable
is obscured (Van Heerde et al., 2005).
Additionally, many online retailers offer free ship-
ping when a customer’s order value reaches a certain
threshold, and such free shipping policies have been
shown to lead to a greater return rate (Lepthien and
Clement, 2019). The study was performed in collab-
oration with a retailer of streetwear and sportswear
items, and the aforementioned results were obtained
using OLS and logistic regressions.
Payment Methods, Assortment Diversity and His-
torical Records: In the Business-to-Business (B2B)
domain, whilst manufacturers have the statistics for
the returns that they receive from retailers’ leftover
stock after the end of a season, they may not necessar-
ily have access to valuable customer returns data from
the retailer, which could in fact help them to reduce the
number of B2B returns. Yan and Cao (2017) used data
shared by an online retailer with the manufacturer of
its products, which consisted of shoes, apparel, and ac-
cessories to gain insights. Firstly, it was found that the
payment method used by the customer was a helpful
predictor of product returns: credit cards, encouraging
a ’buy-now-pay-later’ mindset leading to impulsive
and low-effort purchases, were found to be associated
with a high return rate. On the other hand, paying by
cash was found to curb consumers’ impulsive urges
to make unnecessary purchases, and was associated
with a low return rate. Secondly, if the assortment of
apparel, footwear, and accessory items ordered by the
customer is very diverse results on lower return rates.
Therefore it is very important to distinguish between
bracketing and simply buying many items at once to
keep. Lastly, the return rate was also found to be in-
versely proportional to the number of items that the
customer had ordered from the online retailer in the
past. This study strongly reinforces the importance of
B2C-level information for predicting returns.
Our literature search did not identify any AI or
non-AI based approach that addresses a recommender
system to provide guidance on how best to handle
returned items. Predicting returns provides the basis
for handling returned items in the best possible way
and tailoring reverse logistics processes accordingly.
3.2 On Reverse Logistics Network
Design and Optimisation
Reverse logistics is defined as ”the process of moving
goods from their typical final destination for the pur-
pose of capturing value or ensuring proper disposal”
(Chileshe et al., 2016). Reverse logistics planning is
ICSBT 2023 - 20th International Conference on Smart Business Technologies
18

Table 1: Overview of literature for each research topic, sorted by subcategory and approach.
Topic Subcategory Approaches Literature
Consumer
Returns
Forecasting
Comparison of
Machine
Learning
Methods
Tree-based models, lazy, rule-based, func-
tional algorithms
T
¨
uyl
¨
u and Ero
˘
glu (2019)
Data mining methods Asdecker and Karl (2018)
Mahalanobis feature extraction Urbanke et al. (2015)
Time-series based model Shang et al. (2020)
Observations
of Consumer
Behavior
Subject to demographics, payment method
preference, bracketing
Makkonen et al. (2021); As-
decker et al. (2017)
Effect of pricing, effect of accuracy and
availability of product reviews
Sahoo et al. (2018)
Free shipping promotions, threshold-based
free shipping policies
Shehu et al. (2020); Lep-
thien and Clement (2019)
Effects of payment methods, assortment di-
versity, and past customer records
Yan and Cao (2017)
Reverse
Logistics
Network
Design
Digital Reverse
Logistics Twin
Real-time data, smart robots and AI-based
data predictions (e.g. for vehicle routing)
Ivanov and Dolgui (2021);
Zhang et al. (2018); Sun
et al. (2022a,b)
Mathematical
Programming
Mixed linear programming (MILP) (e.g. fa-
cility location-allocation)
Das et al. (2020)
Hybrid multicriteria decision-making
(MCDM) (e.g. selection of 3PRLP)
Wang et al. (2021)
Collection
Machine Learn-
ing
Hierarchical Clustering for facility location-
allocation
Lin et al. (2021);
Nanayakkara et al. (2022)
Mathematical
Programming
Fuzzy MCDM for collection optimisation
Sagnak et al. (2021);
Ocampo et al. (2019)
Fuzzy MONLP green vehicle routing Soleimani et al. (2017)
MINLP facility location-allocation Vahdani et al. (2012)
Warehousing
AI- and
Vision-based
Systems
AI- and vision-based identification, inspec-
tion and sorting in combination with smart
robots
Wang et al. (2020); Zhang
et al. (2018)
Cobots and AI-based systems for automatic
sorting
Sarc et al. (2019)
Processing
Remanufacturing
RFID, Chaos-based Interactive Artificial
Bee Colony (CI-ABC)
Kumar et al. (2015)
Textile
Recycling
Artificial Neural Networks (ANN) Furferi and Governi (2008)
Zero-waste, second hand clothing Lewis et al. (2017)
Design for improved recycling Durham et al. (2015)
Textile to textile: drivers & inhibiters Sandvik and Stubbs (2019)
Recycling and reuse of clothing in general Xie et al. (2021)
more difficult than forward logistics because there is
great uncertainty about the quantity, timing and quality
of the returned products (Flapper, 1995). The ability
to more accurately predict the timing and volume of
returns using AI-based predictions is essential for op-
timising the reverse logistics process. This enables
optimal performance in terms of collection, transporta-
tion, remanufacturing and recycling (Tibben-Lembke
Advances in AI-Based Garment Returns Prediction and Processing: A Conceptual Approach for an AI-Based Recommender System
19

and Rogers, 2002). Forecasting returns is therefore im-
portant for the design of the reverse logistics network
and the planning and control of the recovery processes
(Xiaofeng and Tijun, 2009). Based on Agrawal et al.
(2015) and Wilson et al. (2021), this research paper
distinguishes four main activities within the reverse
logistics process that are also relevant to the fashion
and apparel industry: network design, collection, ware-
housing and processing. In the context of the follow-
ing, this research paper will examine the contribution
that AI makes, or can make, to the four components of
reverse logistics.
3.2.1 Network Design in Reverse Logistics
Network design is a strategic consideration for all re-
verse logistics functions. Managers need to decide at
a strategic level how the reverse logistics infrastruc-
ture should be structured in general. This includes, for
example, the number and geographical location of lo-
gistics sites and the organisation of transport between
them. There are many differences between reverse and
forward logistics, which is why reverse logistics can-
not simply copy the process of forward logistics. When
goods are returned to the manufacturer, they are sent
from many, initially unknown, points to a central loca-
tion and vary greatly in quality (Wilson et al., 2021).
This lack of visibility and apparent unpredictability
can be addressed with AI-based and non-AI-based
approaches discussed in the previous sections of this
paper.
Using AI in Network Design: AI and real-time data
integration are often seen as the basis for making
strategic, tactical and operational decisions regarding a
smart and sustainable reverse logistics network design
(Sun et al., 2022b). In combination with mathemati-
cal models and computer-aided simulations, AI-based
data predictions and real-time data can be used to build
a digital reverse logistics twin in the sense of Reverse
Logistics 4.0 (Ivanov and Dolgui, 2021). This can then
be used to make informed decisions about the design
of the reverse logistics network by reducing uncertain-
ties and, for example, planning optimal scheduling
and vehicle routing (Zhang et al., 2018). Reasonable
decision parameters for vehicle routing could be, for
example, economic costs, truck utilisation, greenhouse
gas reduction and truck driver working hours (Sun
et al., 2022b). Developing an intelligent digital reverse
logistics twin to inform network design decisions re-
quires deep methodological integration and system
integration, including smart robots and devices, ana-
lytical models, visualisation tools, etc., which must be
effectively and seamlessly linked (Sun et al., 2022a).
Mathematical Methods and Network Design: Ac-
cording to the current state of research, the selection
of third-party reverse logistics providers (3PRLP) is
not based on AI, but rather on mathematical models.
Wang et al. (2021) use a hybrid multicriteria decision
making (MCDM) approach (FAHP combined with
FTOPSIS) to improve 3PRLP evaluation and selec-
tion practices for several industries and to address the
increasing demand for outsourcing reverse logistics
activities. In their empirical case study focusing on
a fashion retailer, Das et al. (2020) investigate which
locations are best suited for initial collection centres
(ICCs), where customer returns are stored for some
time before being sent to final warehouses. They used
Mixed Linear Programming (MILP) as an approach to
minimise the costs (environmental and economic) of
the reverse logistics network.
3.2.2 Collection Approaches Including AI
Methods
For the creation of a sustainable reverse logistics net-
work, it is essential to optimise the routes the returned
goods take to their next destination to either be re-
paired, resold, reused, remanufactured or recycled
(Alkahtani et al., 2021). This gives rise to the ques-
tion of how and where to collect the goods, which
in many cases can be answered using mathematical
optimisation models and Machine Learning. In re-
cent years, Artificial Intelligence models have been
developed in order to optimise the collection process
(Wilson et al., 2021). Additionally, smart collection
techniques emerged in recent years, utilising industry
4.0 technologies like IoT, big data, cloud technology,
virtual technology, autonomous robots or Artificial
Intelligence to make the collection process more re-
source efficient (Sun et al., 2022b).
Fuzzy Multi Criteria Decision Making Methods:
A popular approach considering the collection of prod-
ucts, especially End-of-Life (EoL) products, is the
utilisation of fuzzy logic and probabilistic models to
make up for uncertainties regarding time, quantity, in-
volved parties and more. The selection of locations
for collection centres is thereby often optimised by
multi criteria decision making (MCDM), which allows
for the consideration of several, partially contradicting
and complex criteria (Sagnak et al., 2021). MCDM
methods like analytic network process (ANP), analytic
hierarchy process (AHP), Best-Worst, DEMATEL or
TOPSIS are combined with fuzzy set theory (FST)
and fuzzy logic for improving the product disposition
process (Ocampo et al., 2019; Sagnak et al., 2021).
ICSBT 2023 - 20th International Conference on Smart Business Technologies
20

Mixed-Integer Linear and Non-Linear Program-
ming: Soleimani et al. (2017) investigate a multi-
objective non-linear programming model for the green
vehicle routing problem (GVRP) including pickup
of EOL products using a fuzzy approach. Vahdani
et al. (2012) present a possibilistic mixed non-linear
programming model for facility location-allocation in
closed loop supply chain network design, which shows
efficiency improvements. Yang and Chen (2020)
tackle the collection centre location problem using a
multi-objective mixed integer linear programming and
fuzzy ANP approach to improve the decision-making
process.
Hierarchical Clustering: Another approach to
solve the location allocation problem subject to trans-
portation efficiency and cost of (collection) facilities
is Hierarchical Clustering, an unsupervised Machine
Learning technique, exploring the cluster patterns of
data (Lin et al., 2021). Location allocation problems
have also been solved by supervised Machine Learn-
ing techniques such as k-means clustering, applied
to contexts other than collection centres (Zhou et al.,
2020). Lin et al. (2021) provide a hierarchical clus-
tering framework to optimise logistics processes by
selecting facility nodes of logistics networks including
warehouses, distribution centres and terminal stations,
from a given set of options. Nanayakkara et al. (2022)
develop a method optimising either the collection cen-
tre locations or which geographical areas are assigned
to preexisting collection centres, using a three-step
approach. First, ward-like hierarchical clustering with
geographical constraints is applied, followed by the
determination of the best location of initial collection
centres (ICC), or selection of the best preexisting ICCs,
for each cluster based on a centre of gravity calculation.
From this point on, a network design is created and
optimised regarding sustainability and other factors
(Nanayakkara et al., 2022).
3.2.3 Warehousing Approaches Including AI
Methods
Once the returned parts have been collected, the ware-
housing process begins as the next step in reverse lo-
gistics. This includes tasks such as inspection, sort-
ing, consolidation and inventory management (Wilson
et al., 2021). Typically, the condition of returned prod-
ucts is unknown and can vary widely. This makes
inspection an essential and labour-intensive step in
the warehousing process (Bai and Sarkis, 2013). The
number of returns is also unknown, so the approaches
described earlier in this study for predicting product
returns play an important role in the planning of the
warehousing process. This allows, for example, ca-
pacity to be planned and inventory management activi-
ties to be coordinated at an early stage (Wilson et al.,
2021).
AI- and Vision-Based Systems: Sorting is an im-
portant step in the warehousing process as it is here
that decisions are made on what to do with returned
fashion products once they have arrived at the ware-
house and been inspected. Based on the evaluation of
the returned items, a decision can be made whether
the garments should be reused, repaired, refurbished
or recycled, all of which are steps in what is known
as ’processing’, which is described below. The aim
of warehousing activities is to get returned items back
into use as quickly as possible. To this end, the items
are stored in consolidated form. Inventory activities
such as counting, tracking, sorting are therefore com-
mon tasks in reverse logistics warehousing (Wilson
et al., 2021). AI and vision-based systems can, for ex-
ample, enable smart robots to recognise the different
types of recyclable materials and sort them accordingly
(Wang et al., 2020; Zhang et al., 2019).
Cobots and AI-Based Assistance Systems: Auto-
mated sorting often focuses on ensuring that workers
do not come into contact with hazardous materials.
For this reason, collaborative robots (’cobots’) are in-
creasingly being used, for example to complete a task
started by a human worker (Sarc et al., 2019). But
even in the context of the fashion and apparel supply
chain, where hazardous components play a minor role
in reverse logistics, it is important that decisions about
reprocessing, reuse or recycling are made quickly in or-
der to be as economical and environmentally as possi-
ble. To the best of our knowledge, there is no approach
yet that offers an AI-based assistance system in the
context of reverse logistics in the fashion and apparel
sector to help organise the warehousing process.
3.2.4 Processing Approaches, Including AI
Methods
The final step in reverse logistics is the processing of
the returned goods depending on their condition. In
a reverse logistics network as a part of a closed-loop
supply chain, it is crucial to find an alternative to the
destruction of garments or final disposal in landfills.
Decision Systems: In a circular economy model,
subsequent to inspection and sorting, the goods are
individually processed according to the best option
available. Options usually include reuse, repair, reman-
ufacturing, recycling or disposal (Wilson et al., 2021).
Advances in AI-Based Garment Returns Prediction and Processing: A Conceptual Approach for an AI-Based Recommender System
21

For that matter, Abdessalem et al. (2012) propose de-
cision modeling to find the best reprocessing option
for any EoL return and apply this technique in two
industrial cases. Disassembly forms an essential part
of the processing in reverse logistics and a candidate
for AI applications (Wilson et al., 2021), but is rarely
investigated for the fashion industry. Shahidzadeh and
Shokouhyar (2022) performed a social media analysis,
employing Convolutional Neural Networks (CNNs)
and Long short-term memory (LSTM) to achieve
consumer-centric disposition decision support for man-
agers in RL (Shahidzadeh and Shokouhyar, 2022). For
instance, extracted happiness spectra from tweets re-
veal the contentment of consumers with specific fea-
tures, which is mapped onto one out of three decisions
(refurbish, repair and reuse, recycling) (Shahidzadeh
and Shokouhyar, 2022). From this approach, bench-
marks for developing and developed countries are de-
rived (Shahidzadeh and Shokouhyar, 2022). If in suf-
ficiently good condition, goods such as garments can
be directly resold in the primary market or in a sec-
ondary market. The most suitable market needs to
be selected and the goods are integrated back into the
forward supply chain. While there exist many rec-
ommender systems for apparel on the consumer side
(Mohammed Abdulla et al., 2019; Kottage et al., 2018;
Bellini et al., 2022), our literature review failed to find
any AI- or non AI-based recommender system for find-
ing the most suitable sales channel for returned items
based on their properties.
Reusing, Remanufacturing and Recycling: When
it comes to remanufacturing, using IoT technologies
like RFID form one approach to improve the process.
Kumar et al. (2015) propose an Chaos-based Interac-
tive Artificial Bee Colony (CI-ABC) algorithm to test
the effect of RFID in reverse logistics. The implemen-
tation of RFID leads to substantially increased overall
costs due to the investments in equipment, however,
the operational time performance increases more sub-
stantially. Fabric and apparel recycling strategies have
been investigated in recent years (Xie et al., 2021).
Lewis et al. (2017) propose a zero-waste model for
second-hand apparel by finding repurposing strategies.
Payne (2015) investigates open- and closed-loop recy-
cling methods of textile products, also in the context
of fashion. However, recycling techniques for textiles
are still limited due to fragmentation of supply chains
and limited technology, especially in the field of ma-
terial separation (Sandvik and Stubbs, 2019). A new
potential therefore lies in innovative materials as well
as improved collection and enhanced collaboration
among stakeholders (Sandvik and Stubbs, 2019). The
first of the mentioned potentials has been tackled by
Durham et al. (2015), who discuss important aspects
to be considered in the design process, which can lead
to improved recyclability at the end of the apparel life
cycle. Furferi and Governi (2008) employ a matrix
approach combined with a self-organising feature map
(SOFM) and a feed-forward backpropagation artifi-
cial neural network (FFBP ANN)-based approach for
colour classification of wool-clothing, also taking into
account the respective recycling process and colour
similarity to get an optimal colour and material when
merging two wool-materials for recycling. In the scope
of this paper, literature findings about recycling and re-
manufacturing of apparel are very limited, but relevant
for future research.
All findings from the literature review above are
summarised in Table 1.
4 AI-BASED RECOMMENDER
SYSTEM
The identified research gaps in the area of data-driven
and partially automated analysis of customer returns
data are to be closed in the underlying research project,
which is currently being conducted in Germany. Based
on the findings of the state of the art, a conceptual
approach for an AI-supported recommender system
targeting the prediction of customer returns and the op-
timised handling of follow-up processes in the fashion
and apparel industry was developed. Interviews and
workshops with experts and stakeholders in the indus-
try were conducted to ensure the application proximity
and user-centric focus of the concept. The main com-
ponents of the recommender system are explained in
more detail below.
4.1 Prediction of Future Returns
The system component enables the user to view re-
turn probabilities at the item level. Machine Learning
models are used to continuously predict the potential
return based on product information, customer’s previ-
ous orders, return history, shopping cart information
combined with the payment method and the elapsed
time since order as input variables. In addition, the
reason for the return, such as quality related returns,
bracketing or false item returns, will be determined.
The system thus takes into account both customer-
related and product-related return aspects. To train the
Machine Learning models supervised, existing return
data and completed orders from the systems are used,
which include the ground truth variables of the return
reason. Additionally, the Machine Learning results of
all open orders will be aggregated and presented in a
ICSBT 2023 - 20th International Conference on Smart Business Technologies
22

user-friendly dashboard. During the application of the
system, the Machine Learning models are retrained
cyclically based on the newly generated ground truth
data. Thanks to the available and aggregated informa-
tion, returns management processes can be triggered
and planned earlier. Among other things, this leads to
an acceleration of the processes while at the same time
enabling targeted forecasts of future return costs.
4.2 Decision Assistance for a Sustainable
Second Life Planning of Returns
Usually, returned items are assigned to appropriate
follow-up processes based on the reason for return
and an initial assessment. Possible processing steps
include additional quality checks, repairs or cleaning.
Finally, the goods are destroyed, recycled or sold as
’new’ in a secondary channel. Decisions for the follow-
up processes of returns have so far mostly been made
manually and thus are prone to be subjectively influ-
enced. In the second system component, all available
information is used to recommend decisions on the
follow-up process at a product level. With information
about, for example, the duration of the return process,
the reason for the return, the result of the first inspec-
tion, the product group and product type, an AI model
is trained to issue a recommendation to the user on
how to proceed with the returned goods and via which
channel the item should be offered. The system is
combined with developed metrics from sustainability
analyses and expert information to additionally incor-
porate the sustainability influences resulting from the
processes in the decision-making basis. By improv-
ing the quality of decisions and reducing process time,
the system aims to optimise the process and reduce
economic losses. Additionally, it aims to ensure that
destruction or recycling of returned goods are only con-
sidered as the last option. This reduces the waste of
material, the resulting carbon footprint and contributes
to the circular economy in the fashion and apparel
sector.
5 CONCLUSION AND FUTURE
WORK
The fashion and apparel sector is responsible for the
majority of returns in e-commerce. In the context
of the present research, it could be shown that there
are a large number of studies dealing with the topic
of reverse logistics network design, but these studies
mostly refer to the industrial sector, e.g. in the sense
of remanufacturing. It is difficult to find research that
deals specifically with the returns process of apparel
products in the e-commerce sector. Some transfer of
studies from other sectors is possible, but the large
number of variants in the fashion and apparel sector
and the sheer volume of returns pose a particular chal-
lenge for manufacturers and retailers. Due to the high
environmental impact of the industry, additional re-
search is needed, which should primarily deal with the
most sustainable processing of returned articles and the
optimisation potential for this in the context of reverse
logistics processes. The paper provides an overview
of relevant developments in the fashion and apparel
industry and does not claim to be a systematic and
complete review. However, given the environmental
and economic relevance of the issue, this should also
be the aim of further studies. Furthermore, it could
be shown that although there are approaches to use
AI and ML in the field of returns forecasting and re-
verse logistics network design, mathematical methods
dominate and their suitability to the underlying com-
plexity is limited due to the mostly static approaches.
In addition, there are currently no appropriate AI and
ML applications that can be integrated into a system
that is used in practice (e.g. ERP or PDM). For these
reasons, it is advisable to focus more on approaches
and applications in the context of AI and ML in the
future. The knowledge gained from this research will
be used as a guideline for the design of an AI-based
recommender system that will provide meaningful rec-
ommendations for the further processing of returns
based on returns predictions, thus making a valuable
contribution from an environmental and economic per-
spective in an industry that is responsible for heavy
pollution.
ACKNOWLEDGEMENTS
This research was funded in part by the Ger-
man Federal Ministry of Education and Research
(BMBF) under the project OptiRetouren (grant num-
ber 01IS22046B). It is a joint project of the August-
Wilhelm Scheer Institut, INTEX, HAIX and h+p.
August-Wilhelm Scheer Institut is mainly entrusted
with conducting research in AI for forecasting returns
volume and for recommendations based on AI.
REFERENCES
Abdessalem, M., Hadj-Alouane, A. B., and Riopel, D.
(2012). Decision modelling of reverse logistics sys-
tems: selection of recovery operations for end-of-life
Advances in AI-Based Garment Returns Prediction and Processing: A Conceptual Approach for an AI-Based Recommender System
23

products. International Journal of Logistics Systems
and Management, 13(2):139–161.
Agrawal, S., Singh, R. K., and Murtaza, Q. (2015). A lit-
erature review and perspectives in reverse logistics.
Resources, Conservation and Recycling, 97:76–92.
Alkahtani, M., Ziout, A., Salah, B., Alatefi, M., Abd El-
gawad, A. E. E., Badwelan, A., and Syarif, U. (2021).
An insight into reverse logistics with a focus on collec-
tion systems. Sustainability, 13(2):548.
Asdecker, B. and Karl, D. (2018). Big data analytics in
returns management-are complex techniques necessary
to forecast consumer returns properly? In 2nd Interna-
tional Conference on Advanced Research Methods and
Analytics. Proceedings, pages 39–46.
Asdecker, B., Karl, D., and Sucky, E. (2017). Examining
drivers of consumer returns in e-tailing with real shop
data. In Hawaii International Conference on System
Sciences, pages 4192–4201.
Bai, C. and Sarkis, J. (2013). Flexibility in reverse logis-
tics: a framework and evaluation approach. Journal of
Cleaner Production, 47:306–318.
Bellini, P., Palesi, L. A. I., Nesi, P., and Pantaleo, G. (2022).
Multi clustering recommendation system for fashion
retail. Multimedia Tools and Applications, pages 1–28.
Bimschleger, C., Patel, K., and Leddy, M. (2019). Bringing
it back: Retailers need a synchronized reverse logistics
strategy. Technical report, Deloitte Development LLC.
Chileshe, N., Rameezdeen, R., and Hosseini, M. R. (2016).
Drivers for adopting reverse logistics in the construc-
tion industry: a qualitative study. Engineering, Con-
struction and Architectural Management, 23(2):134–
157.
Das, D., Kumar, R., and Rajak, M. (2020). Designing a
reverse logistics network for an e-commerce firm: A
case study. Operations and Supply Chain Management:
An International Journal, 13(1):48–63.
Deges, F. (2021). Retourencontrolling im online-handel.
Controlling – Zeitschrift f
¨
ur erfolgsorientierte Un-
ternehmenssteuerung, 2/2021:61–68.
Durham, E., Hewitt, A., Bell, R., and Russell, S. (2015).
Technical design for recycling of clothing. In Sustain-
able apparel, pages 187–198. Elsevier.
Flapper, S. D. P. (1995). One-way or reusable distribution
items? TU Eindhoven. Fac. TBDK, Vakgroep LBS:
working paper series, 9504.
Forschungsgruppe Retourenmanagement (2022).
Ergebnisse des europ
¨
aischen retourentachos
ver
¨
offentlicht. https://www.retourenforschung.de/info-
ergebnisse-des-europaeischen-retourentachos-
veroeffentlicht.html. Online; accessed 2023-01-26.
Furferi, R. and Governi, L. (2008). The recycling of wool
clothes: an artificial neural network colour classifi-
cation tool. The International Journal of Advanced
Manufacturing Technology, 37:722–731.
Heckman, J. J. (1979). Sample selection bias as a specifica-
tion error. Econometrica: Journal of the econometric
society, pages 153–161.
Holmes, G., Hall, M., and Prank, E. (1999). Generating rule
sets from model trees. In Australasian joint conference
on artificial intelligence, pages 1–12. Springer.
Ivanov, D. and Dolgui, A. (2021). A digital supply chain
twin for managing the disruption risks and resilience in
the era of industry 4.0. Production Planning & Control,
32(9):775–788.
Jenkins, G. M. (1970). Time Series Analysis; Forecasting
and Control [by] George EP Box and Gwilym M. Jenk-
ins. San Francisco: Holden-Day.
Kottage, G. N., Jayathilake, D. K., Chankuma, K. C., Gane-
goda, G. U., and Sandanayake, T. (2018). Preference
based recommendation system for apparel e-commerce
sites. In 17th international conference on computer
and information science, pages 122–127. IEEE.
Kumar, V. V., Liou, F. W., Balakrishnan, S., and Kumar, V.
(2015). Economical impact of rfid implementation in
remanufacturing: a chaos-based interactive artificial
bee colony approach. Journal of Intelligent Manufac-
turing, 26:815–830.
Lepthien, A. and Clement, M. (2019). Shipping fee sched-
ules and return behavior. Marketing Letters, 30(2):151–
165.
Lewis, T. L., Park, H., Netravali, A. N., and Trejo, H. X.
(2017). Closing the loop: A scalable zero-waste model
for apparel reuse and recycling. International Jour-
nal of Fashion Design, Technology and Education,
10(3):353–362.
Lickert, H., Wewer, A., Dittmann, S., Bilge, P., and Diet-
rich, F. (2021). Selection of suitable machine learning
algorithms for classification tasks in reverse logistics.
Procedia CIRP, 96:272–277.
Lin, T., Liu, Y., Liu, B., Wang, Y., Wu, S., and Zhe, W.
(2021). Hierarchical clustering framework for facility
location selection with practical constraints. IET Cyber-
Physical Systems: Theory & Applications, 6(4):238–
253.
Makkonen, M., Frank, L., and Kemppainen, T. (2021).
The effects of consumer demographics and payment
method preference on product return frequency and rea-
sons in online shopping. In Bled eConference, pages
567–580. University of Maribor.
Mohammed Abdulla, G., Singh, S., and Borar, S. (2019).
Shop your right size: A system for recommending sizes
for fashion products. In Companion Proceedings of
The 2019 World Wide Web Conference, pages 327–334.
Morgan Stanley (2022). Here’s why e-commerce growth can
stay stronger for longer. https://www.morganstanley.
com/ideas/global-ecommerce-growth-forecast-2022/.
Online; accessed 2023-01-26.
Nanayakkara, P. R., Jayalath, M. M., Thibbotuwawa, A.,
and Perera, H. N. (2022). A circular reverse logistics
framework for handling e-commerce returns. Cleaner
Logistics and Supply Chain, 5:100080.
Ocampo, L., Himang, C., Kumar, A., and Brezocnik, M.
(2019). A novel multiple criteria decision-making ap-
proach based on fuzzy dematel, fuzzy anp and fuzzy
ahp for mapping collection and distribution centers in
reverse logistics. Advances in Production Engineering
& Management, 14(3):297–322.
Payne, A. (2015). Open-and closed-loop recycling of textile
and apparel products. In Handbook of life cycle assess-
ICSBT 2023 - 20th International Conference on Smart Business Technologies
24

ment (LCA) of textiles and clothing, pages 103–123.
Elsevier.
Quinlan, J. R. et al. (1992). Learning with continuous classes.
In 5th Australian joint conference on artificial intelli-
gence, volume 92, pages 343–348. World Scientific.
Sagnak, M., Berberoglu, Y., Memis,
˙
I., and Yazgan, O.
(2021). Sustainable collection center location selec-
tion in emerging economy for electronic waste with
fuzzy best-worst and fuzzy topsis. Waste Management,
127:37–47.
Sahoo, N., Dellarocas, C., and Srinivasan, S. (2018). The
impact of online product reviews on product returns.
Information Systems Research, 29(3):723–738.
Sandvik, I. M. and Stubbs, W. (2019). Circular fashion sup-
ply chain through textile-to-textile recycling. Journal
of Fashion Marketing and Management: An Interna-
tional Journal, 23(3):366–381.
Sarc, R., Curtis, A., Kandlbauer, L., Khodier, K., Lorber,
K. E., and Pomberger, R. (2019). Digitalisation and
intelligent robotics in value chain of circular economy
oriented waste management–a review. Waste Manage-
ment, 95:476–492.
Schwaiger, R. and Steinwendner, J. (2019). Neuronale Netze
programmieren mit Python. Rheinwerk Computing.
Shahidzadeh, M. H. and Shokouhyar, S. (2022). Shedding
light on the reverse logistics’ decision-making: a social-
media analytics study of the electronics industry in de-
veloping vs developed countries. International Journal
of Sustainable Engineering, 15(1):161–176.
Shang, G., McKie, E. C., Ferguson, M. E., and Galbreth,
M. R. (2020). Using transactions data to improve con-
sumer returns forecasting. Journal of Operations Man-
agement, 66(3):326–348.
Shehu, E., Papies, D., and Neslin, S. A. (2020). Free shipping
promotions and product returns. Journal of Marketing
Research, 57(4):640–658.
Soleimani, H., Govindan, K., Saghafi, H., and Jafari, H.
(2017). Fuzzy multi-objective sustainable and green
closed-loop supply chain network design. Computers
& industrial engineering, 109:191–203.
Sun, X., Yu, H., and Solvang, W. D. (2022a). System in-
tegration for smart reverse logistics management. In
2022 IEEE/SICE International Symposium on System
Integration (SII), pages 821–826. IEEE.
Sun, X., Yu, H., and Solvang, W. D. (2022b). Towards
the smart and sustainable transformation of reverse
logistics 4.0: a conceptualization and research agenda.
Environmental Science and Pollution Research, pages
1–19.
Tibben-Lembke, R. S. and Rogers, D. S. (2002). Differences
between forward and reverse logistics in a retail envi-
ronment. Supply Chain Management: An International
Journal, 7(5).
T
¨
uyl
¨
u, A. N. A. and Ero
˘
glu, E. (2019). Using machine learn-
ing algorithms for forecasting rate of return product
in reverse logistics process. Alphanumeric Journal,
7(1):143–156.
Urbanke, P., Kranz, J., and Kolbe, L. M. (2015). Predict-
ing product returns in e-commerce: The contribution
of mahalanobis feature extraction. In International
Conference on Interaction Sciences, pages 1–19.
Vahdani, B., Razmi, J., and Tavakkoli-Moghaddam, R.
(2012). Fuzzy possibilistic modeling for closed loop re-
cycling collection networks. Environmental Modeling
& Assessment, 17:623–637.
Van Heerde, H. J., Gijsbrechts, E., and Pauwels, K. (2005).
Price war: what is it good for? store incidence and
basket size response to the price war in dutch grocery
retailing. Tilburg University, LE Tilburg, The Nether-
lands.
Vogue/BCG (2021). Consumers’ adaption to sus-
tainability in fashion. https://web-assets.bcg.
com/27/f3/794284e7437d99a71d625caf589f/
consumers-adaptation-to-sustainability-in-fashion.
pdf. Online; accessed 2023-01-26.
Wang, C.-N., Dang, T.-T., and Nguyen, N.-A.-T. (2021).
Outsourcing reverse logistics for e-commerce retailers:
A two-stage fuzzy optimization approach. Axioms,
10(1):34.
Wang, Z., Li, H., and Yang, X. (2020). Vision-based robotic
system for on-site construction and demolition waste
sorting and recycling. Journal of Building Engineering,
32:101769.
Wilson, M., Paschen, J., and Pitt, L. (2021). The circular
economy meets artificial intelligence (ai): Understand-
ing the opportunities of ai for reverse logistics. Man-
agement of Environmental Quality: An International
Journal, 33(1):9–25.
Xiaofeng, X. and Tijun, F. (2009). Forecast for the amount
of returned products based on wave function. In 2009
International Conference on Information Management,
Innovation Management and Industrial Engineering,
volume 2, pages 324–327. IEEE.
Xie, X., Hong, Y., Zeng, X., Dai, X., and Wagner, M. (2021).
A systematic literature review for the recycling and
reuse of wasted clothing. Sustainability, 13(24):13732.
Yan, R. and Cao, Z. (2017). Product returns, asymmetric
information, and firm performance. International Jour-
nal of Production Economics, 185:211–222.
Yang, C. and Chen, X. (2020). A novel approach integrat-
ing fanp and momilp for the collection centre location
problem in closed-loop supply chain. International
Journal of Sustainable Engineering, 13(3):171–183.
Zhang, Y., Liu, S., Liu, Y., Yang, H., Li, M., Huisingh, D.,
and Wang, L. (2018). The ‘internet of things’ enabled
real-time scheduling for remanufacturing of automo-
bile engines. Journal of cleaner production, 185:562–
575.
Zhang, Z., Wang, H., Song, H., Zhang, S., and Zhang, J.
(2019). Industrial robot sorting system for municipal
solid waste. In Intelligent Robotics and Applications:
12th International Conference, Proceedings, Part II 12,
pages 342–353. Springer.
Zhou, Y., Xie, R., Zhang, T., and Holguin-Veras, J. (2020).
Joint distribution center location problem for restaurant
industry based on improved k-means algorithm with
penalty. IEEE Access, 8:37746–37755.
Advances in AI-Based Garment Returns Prediction and Processing: A Conceptual Approach for an AI-Based Recommender System
25