Towards Minimizing e-Commerce Returns for Clothing
A. K. Seewald
1
, T. Wernbacher
2
, A. Pfeiffer
2
, N. Denk
2
, M. Platzer
3
, M. Berger
3
and T. Winter
4
1
Seewald Solutions, L
¨
archenstraße 1, A-4616 Weißkirchen a.d. Traun, Austria
2
Donau-Universit
¨
at Krems, Dr.-Karl-Dorrek-Straße 30, A-3500 Krems, Austria
3
yVerkehrsplanung, Brockmanngasse 55, A-8010 Graz, Austria
4
Attribu-i, Nibelungengasse 32d, A-8010 Graz, Austria
Keywords:
Machine Learning, Visualization, Data Mining, Rule Learning, e-Commerce, Returned Goods.
Abstract:
The importance of e-commerce including the associated freight traffic with all its negative consequences (e.g.
congestion, noise, emissions) is constantly increasing. Already in 2015, an European market volume of 444
billion Euros at an annual growth of 13.3% was achieved, of which clothing and footwear account for 12.7%
as the largest category (Willemsen et al., 2016). However, online commerce will only have a better footprint
than buying in the local retail shop under optimal conditions (for example: group orders, always present at
home delivery, no returns and no same day delivery). Next to frequent single deliveries, CO
2
intensive and
underutilized transport systems, returned goods are the main problem of online shopping. The last is cur-
rently estimated at up to 50% (Hofacker and Langenberg, 2015; Kristensen et al., 2013). Our research project
Think!First tackles these problems in freight mobility by using an unique combination of gamification ele-
ments, persuasive design principles and machine learning. Customers are animated, targeted and nudged to
choose effective and sustainable means of transport when shopping online while ensuring best fit by compen-
sating both manufacturer and customer biases in body size estimation. Here we show preliminary results and
also present a slightly modified rule learning algorithm that always characterizes a given class (here: returns).
1 INTRODUCTION
e-Commerce is nationally and internationally on the
rise (Knabl et al., 2015). Especially retail e-commerce
in clothes and shoes over the internet – just for Europe
– was already at 56 billion EUR in 2015 and is rising
at a yearly rate of 13.3%, forecasted to double until
2020 (Willemsen et al., 2016).
In Austria the market volume in 2014 was 2.1 bil-
lion EUR corresponding to a growth of 11.6% from
2013 (Knabl et al., 2015). The ten biggest retailers
in Austria received 46% of their sales volume from
online shops (Hofacker and Langenberg, 2015) and
an increasing number of local shops aim to follow
their lead. This increase in online shops is also re-
flected in the shopping behaviour of Austrian cus-
tomers. In 2013 already 57% of Austrian customers
bought items over the internet with a total sales vol-
ume of 5.9 billion EUR spent in Austria and abroad
(Lengauer et al., 2015). Because of the similarities
with the global situation and the early-adopter sta-
tus w.r.t. internet shopping of Austrian customers we
believe that our study and its conclusions – although
mainly based on data from a large Austrian store and
online shopping chain (one of our project partners) –
are also valid in a global setting.
One important issue in retail e-commerce is the
inherently high returns rate of up to 50% (Hofacker
and Langenberg, 2015; Kristensen et al., 2013). Com-
bined with a higher number of offers for shorter de-
livery times this results in a corresponding increase of
freight traffic and therefore CO
2
emissions at a time
when a reduction of these emissions is needed. On
the other hand an increasing number of online shop-
pers want to buy in a sustainable way (Hagemann,
2015; Halbach et al., 2015). Actual shopping deci-
sions are however less sustainable due to ignorance,
laziness and missing incentives. Our research project
Think!First aims to optimally inform and motivate on-
line customers through three methods.
1. Creating transparency by visualizing the effects of
customer decisions on climate and environment
and nudging customers into a sustainable direc-
Seewald, A., Wernbacher, T., Pfeiffer, A., Denk, N., Platzer, M., Berger, M. and Winter, T.
Towards Minimizing e-Commerce Returns for Clothing.
DOI: 10.5220/0007568508010808
In Proceedings of the 11th International Conference on Agents and Artificial Intelligence (ICAART 2019), pages 801-808
ISBN: 978-989-758-350-6
Copyright
c
 2019 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved
801

tion by drawing on research in persuasive design
and gamification.
2. Optimize logistics towards more sustainable
transport vehicles (eg. load bicycles).
3. Reduce returns by compensating customer bias
(e.g. misjudgment of correct size) and manu-
facturer bias (inconsistent or even incorrect re-
ported product sizes) on the size matching pro-
cess. Non-fitting garments are a known factor
to strongly drive returns (Kristensen et al., 2013;
Singh, 2015).
Within this paper we will only address the last point.
More details on the remaining points can be found
in Wernbacher et al. (2017) or at the project website
https://www.thinkfirst.blog
1
.
2 RELATED RESEARCH
Kristensen et al. (2013) present TrueFit, a system to
determine precise body measurements which can re-
duce returns by up to 30%. However it requires much
effort by potential customers. TrueFit works by com-
bining extensive information provided by customers
on their height, age, weight as well as a set of previ-
ously bought fitting clothes with manufacturer, model
type and given size to determine best fit. While it
therefore tries to compensate both customer and man-
ufacturer bias, in its present form it ignores body size
temporal drift.
2
Colsen (2013) describes StitchFit, a clothes seller
which uses a combination of machine learning and
styling expertise provided by human fashion stylists.
Here, customers regularily receive fitting garments by
post. The machine learning system provides a list
of fitting garments based on initial customer-provided
body size data and the complete history of returned
garments. The fashion stylist chooses garments from
this list and sends complete ensembles to the cus-
tomer. However, he completely ignores the potential
for unstructured data.
3
Singh (2015) analyses reasons for returns within
Indian online market Flipkart, where mainly wom-
ens’ garments are sold directly by the manufacturers.
Apart from a detailed analysis of returns reasons they
also provide a minimal set of measurements for size
1
Presently all information on the website is only avail-
able in German. We apologize for the inconvenience.
2
I.e. changes in body size over time
3
We are intrigued by the possibility of replicating styling
expertise via deep learning on product images or garment
style graphs... but this is a topic for another paper.
tables to reduce returns.
4
Simply changing the shown
size tables for nine manufacturers according to his
recommendations reduced returns dramatically: An
average reduction of absolute returns rate of 9% was
reported with a maximum of 46% – so the luckiest
manufacturer saw their returns rate halved. He also
provides an analysis of returns reasons due to product
quality issues which are also a major cause for returns
within this online market, albeit less relevant within
our project.
Ghaffari (2011) analyses the quality of customer
decisions depending on the presentation of garments
via product images, and finds that customer skill level
in online shopping determines the optimal type of
product images: Seasoned clothes shoppers obtain
better size estimates from product images of garments
worn by realistic models, while novice shoppers ob-
tain better estimates from product images showing
just the unworn clothes on a flat surface. Therefore
it is preferrable to provide both types of images –
and restrict shown image types to the optimal type if
the skill level of the customer is known. Our partner
provides both types of product images as well as a
novel combination of both: images of products worn
by mannequins where the mannequins are afterwards
digitally removed from each image, and additional
product information (such as size and washing tags)
is digitally added. It would be interesting to test the
effect of this new image type on novice and expert
shoppers.
Seewald (2007) creates and analyzes a model to
predict the response of inactive customers to a postal
mailing. He tests the value of feature subset selection
in improving the model and finds little to no effect
for the tested algorithms (Naive Bayes, Hidden Naive
Bayes and robust logistic regression). In this paper
we even found that arbitrarily removing features with
high predictive value – a type of feature subset selec-
tion reminiscent of adversial examples – may in some
cases be beneficial (see section 5).
Toktay (2003) analyses different models to predict
returns via synthetic data. He differentiates between
modelling via periodical data where only the number
of sold and returned products is known (i.e. where
it is not possible to identify products and determine
exactly which products were returned), and modelling
via individual data on product level (i.e. where such a
identification is feasible). For the second case – which
corresponds to our data – he proposes an Expectation
4
Minimal set of measurements: breast width, waist cir-
cumference, shoulder circumference, sleeve diameter at
3
4
height; provide at least UK, US and EU-Sizes and at least
S,M,L,XL,XXL for simple sizes.
ICAART 2019 - 11th International Conference on Agents and Artificial Intelligence
802

0.0001
0.001
0.01
0.1
1
1 10 100 1000 10000 100000
percentage of total returns
Abs. number of customers
percentage of total returns
Figure 1: Log-log plot of customers vs. percentage of total
returns (top left 1 = 1%).
0.0001
0.001
0.01
0.1
1
10
1 10 100 1000 10000 100000
percentage of total returns
Abs. number of products
percentage of total returns
Figure 2: Log-log plot of products vs. percentage of total
returns (top left 1 = 1%).
Maximization model. No explicit modelling of the
reasons for returns takes place.
3 INITIAL DATA SURVEY
From the data warehouse of our project partner we re-
ceived a set of a few million samples from a time pe-
riod of several years, containing detailed information
on customers, orders, products, deliveries and returns
during this time period. In total there were 312 fea-
tures – 202 numeric features and 110 nominal features
with up to 10 values (avg. 2.75) per feature. Overall
returns over all products (including clothing) within
this time period were within the ranges reported in
the literature.
One trivial way to reduce returns would be to
delist products or ban customers which are respon-
sible for a large proportion of returns. So we first
checked for power-law distributions as these would
indicate that such an approach is feasible. As can be
seen in Figures 1 and 2 both products and customers
show an approximate line in the log-log plot – with
some outliers – and may in first approximation be
considered a power law distribution. However note
that at the highest point one customer is at most re-
ponsible for 0.1% of total returns and one article is at
most reponsible for 1% of total returns so the slope of
the power-law distribution is in both cases quite small.
0
10000
20000
30000
40000
50000
60000
0 20 40 60 80 100
100
1000
10000
100000
1e+06
Number of products returned
Number of products sold
Number of products returned
Total number of products sold
Figure 3: Products sold vs. products returned vs. returns
rate (X axis), grouped by customers. More details see text.
0
2000
4000
6000
8000
10000
12000
14000
16000
0 20 40 60 80 100
10
100
1000
10000
100000
1e+06
Number of products returned
Number of products sold
Number of products returned
Number of products sold
Figure 4: Products sold vs. products returned vs. returns
rate (X axis), grouped by products. More details see text.
Due to aliasing effects the number of returns, num-
ber of products sold and the returns rate are linked
especially for small number of products sold. To al-
low for a fair comparison, we decided to visualize all
three parameters in a single graph. We binned the re-
turns rate in percent into 100 bins at the X axis. For
each bin we computed both the number of products
sold and the number of returns. The X axis there-
fore shows the returns rate in percent between 0% (no
returns) and 100% (only returns). Absolute number
of products returned and absolute number of products
sold are shown on the left and the right Y axis. For the
number of products sold we had to use a logarithmic
scale to enable a comparison.
Figures 3 and 4 show the results. The difference
between both figures is that for Figure 3 we computed
returns rate on customer level (i.e. each customer was
assigned their average returns rate over all their or-
ders) while in Figure 4 we computed returns rate on
products level (i.e. each product was assigned its av-
erage returns rate over all its sales).
Figure 3 shows a simple pattern. Customers with
very low (left in graph) and very high (right in graph)
return rates cause comparably small returns – but
those with very low return rates tend to buy much
more products than those with very high return rates
so that the absolute number of returns is almost the
same. However the highest number of returns is found
in the center at around 40% to 60% returns rate.
Towards Minimizing e-Commerce Returns for Clothing
803

Figure 4 shows a much more interesting pattern.
The bimodal distribution of returns indicates that re-
turns are driven by two approximate normal distribu-
tions (excluding the outliers around 2%): one smaller
peak of those products which sell very well at a low
relative returns rate – which however translates to a
high number of absolute returns – and a larger peak
of products which sell less well at a higher returns
rate. The latter group seems to contribute about six
times the returns as the former group.
We can now ask what would be necessary to re-
duce the returns rate by e.g. 10%, assuming a well-
trained model that can identify customers and prod-
ucts with high returns a priori with high confidence.
5
To achieve this reduction, we would need to ban
0.17% of customers at the cost of selling 4.31% less
products overall; or delist 1.13% of products at the
cost of selling 13.46% less products overall. Both al-
ternatives were deemed unsatisfactory to our project
partner and therefore rejected.
4 APPROACH
For our approach we initially proposed a system that
does not rely on time-consuming and cumbersome
customer self-reporting
6
while still ensuring a pre-
cise quality customer body size and manufacturer size
data. In each case we tested the data provided by
our partner whether it would be feasible to implement
each point.
1. Reconstruct a precise body size from previous or-
ders that were not returned, using as negative set
those orders which were returned. This approach
is complicated by customers buying for multiple
persons (e.g. parents for children, or wife for hus-
bands / vice versa) and of course necessitates pre-
cise manufacturer size information.
Here we evaluated the available data provided by
our partner and found this approach to be feasible.
We may follow it up in the future.
2. A combination of age (reconstructed from birth
date), sociodemographic and other customer data
could potentially be used to create a rough esti-
mate of height, weight and other body parame-
ters. For ground truth data customers could be
measured when buying products in physical stores
of which our partner has several.
A sufficiently large proportion of the customers in
our dataset have a birth date stored, so for these
5
This may be an unreasonable assumption.
6
E.g. the approach described by Kristensen et al. (2013)
age can be reconstructed. However our partner
considered it too costly to measure customers in
the stores so we were not able to build a customer-
based body size model and had to reject this ap-
proach.
3. Unstructured data such as textual comments and
feedback including five-star-ratings and contin-
uous rating values may be useful to deter-
mine manufacturer-dependent and possibly user-
dependent biases in size estimation.
Our partner provides star-ratings and textual com-
ments for logged-in users which are linked to the
product data so it initially looked feasible to build
such a model for manufacturer-dependent biases.
However, the amount of data was insufficient to
build a model of user-dependent biases so for this
purpose we had to reject the approach.
Concerning manufacturer-dependent bias, we
found that provides detailed size information al-
most on the level of garment cuts to their manu-
facturers who are then producing the garments ac-
cording to specifications. An appropriate amount
of samples chosen by international standards in
garment production are then measured to en-
sure compliance with the initial size information.
In this context there is only a single manufac-
turer and therefore no meaningful manufacturer-
dependent biases to analyze. However it also
means that high-quality consistent size informa-
tion was already available which we could use for
evaluation (see subsection 4.1).
We may revisit the use of unstructured data in fu-
ture work, however for this project it was rejected.
Due to manufacturing tolerances the actual size
may differ about ±1 unit from the size reported
on each garment. We have therefore proposed to
produce garments without any size information at
all and measure each piece delivered by the man-
ufacturer, adding the correct size information and
thereby ensuring perfect size information on each
item.
7
However this approach was also deemed to
be too costly by our partner and also had to be
rejected.
7
This approach is similar to the one used in micropro-
cessor production where maximum operating frequency is
automatically measured and each chip tagged accordingly.
Changes in production quality influence maximum operat-
ing frequency in a complex way – especially at the begin-
ning of a new microprocessor generation – and it is deemed
easier to simply measure the produced chips than to model
the process.
ICAART 2019 - 11th International Conference on Agents and Artificial Intelligence
804

4.1 Quantitative Size Information
As we mentioned during the evaluation of the 3rd ap-
proach, we found out that relatively precise size in-
formation is available for most products. Since it is
known that providing better size information leads to
smaller returns rates (Kristensen et al., 2013; Singh,
2015) and the webshop by our partner only provided
a single size table for all products, we followed up
on this by extracting all size tables from the backend
systems and converted them into a format suitable to
be displayed in the webshop. Among the more than
200 different measurements we restricted ourselves to
those found by Singh (2015) to perform best. Evalua-
tion is still ongoing.
4.2 Qualitative Size Information
Another observation was that – additionally to the
quantitative size information just mentioned – hu-
mans communicate size preferences using qualita-
tive concepts related to fit such as figurative, figure-
accentuating, casual and straight. Our partner gra-
ciously provided such style information for several
hundred products which we used to train a model to
consistently determine qualitative fit information di-
rectly from the size tables.
As features we used arithmetic average, standard
deviation, relative standard deviation, median, max
and min of breast width, inner leg length, waist
circumference and hip circumference over all sizes.
For simplicity we used a robust Logistic Regression
model.
The best model had an accuracy of 59.47% –
much improved over the baseline accuracy of 43.13%
– and we are currently working to improve this by
adding unstructured textual data from product de-
scriptions as well as unstructured image data from
product images. Evaluation is ongoing.
5 CHARACTERIZATION OF
RETURNS
As final step we aimed to characterize returns by a
well-known rule learning algorithm, JRip, which is
an open source implementation of RIPPER (Cohen,
1995) within the data mining suite WEKA.
8
We chose
RIPPER for its ability to produce small concise rule
sets that are easy to interpret. We also considered Lo-
gistic Regression but found that too many features had
8
See https://www.cs.waikato.ac.nz/ml/weka/
Table 1: Successive removal of predictive features as esti-
mated by two-fold crossvalidation.
Dataset #F. Prec. Rec. F
0
AUC
tr3 154 0.813 0.821 0.817 0.840
tr4 150 0.866 0.596 0.706 0.790
tr5 149 0.798 0.772 0.785 0.820
tr6 145 0.822 0.862 0.842 0.857
tr7
11
303 0.865 0.898 0.881 0.917
tr8 290 0.784 0.827 0.805 0.830
tr9 283 0.765 0.793 0.779 0.792
tr10 282 0.765 0.794 0.779 0.794
tr11 237 0.768 0.789 0.778 0.789
tr11 17
12
237 0.767 0.775 0.771 0.769
high weight resp. odds ratio, severly impairing inter-
pretability of the trained model.
9
As we had more
than enough samples, we downsampled a 20% sub-
set of the original dataset to 1:1 class distribution
between returns and non-returns and evaluated the
model on the remaining 80% data that was not sub-
sampled plus the non-returns removed from the train-
ing data during subsampling. For initial evaluation we
used the training set with twofold crossvalidation.
One major problem was that the data dictionary
had not been updated for some time. Therefore from
312 features only about a third were actively used and
well-known. To prevent inadvertently using features
that are changed when returns are entered into the sys-
tem – which may give the system an unfair advantage
and yield results that are too good to be true – we
chose two mitigations: 1) training the model on one
set of data and evaluating on a later data warehouse
export (ongoing), 2) successively removing highly
predictive attributes (either by high odds value in lo-
gistic regression or by appearing often in the first 5-10
rules from RIPPER).
10
Table 1 shows the results for
the second mitigation. It can be seen that the removal
of highly predictive attributes does not always reduce
the performance of RIPPER but in some cases even
proves beneficial. The addition of products base data
from tr7 onward clearly proved beneficial at first and
it will turn out that many final rules make use of those
features.
9
However we stil used it to determine candidate attribute
for removal (see later).
10
The final choice of which attributes to remove was
mainly based on these two criteria since for most candidate
attributes little or no feedback on their meaning could be
obtained from our partner. The cut-off was manually cho-
sen for each dataset by visually inspecting the distribution
of odds values and attributes used in top rules by RIPPER.
11
Adding products base data (VKA.∗); modifying RIP-
PER to force rules on returns (see text)
12
tr11 with only samples from 2017 by order date.
Towards Minimizing e-Commerce Returns for Clothing
805

Initially we ran RIPPER on the complete data
from all years containing a few million samples (tr3-
tr11). However, the data warehouse format had been
changed at least three times during the last several
years in which the data was collected causing some
variables’ interpretation to be changed and some new
variables to be added, both causing RIPPER to return
large sets of around a hundred rules to account for the
additional data variance. So we chose to retrain it us-
ing only the latest data from 2017 and 2018 (tr11 17)
containing a few hundred thousand samples. This led
to a small set of eleven rules which predict returns on
the independent test set with a precision of 0.495 and
recall of 0.784 (balanced F-measure: 0.607), compa-
rable to models trained on the whole dataset.
13
One problem with RIPPER was that according to
its internal heuristics its default class
14
was some-
times returns and sometimes non-returns. However
since rules for non-returns are much harder to inter-
pret than rules for returns and we are also much more
interested in the latter, we chose to adapt RIPPER to
force it to use only non-returns as default class and
therefore always output rules that predict returns. The
modified RIPPER was used from tr7 onwards. tr1
and tr2 had minor errors in preprocessing and were
therefore removed from the table.
We will now interpret the rules from the rule list
in order. Note that class=1 corresponds to samples
with class returns and that it is necessary to apply
these rules in exactly the given order to get correct
results.
(VKA.ARSeit_month >= 4) and (VKA.VK_PreisA >=
49.9) and (VKA.WarennummerCode <=
62069090) => class=1 (7573.0/1469.0)
This rule only utilizes products base data (VKA.∗).
All products which have been in the online shop
since April (ARSeit month – primarily excluding
some products only available in winter) and the sales
price (VK PreisA) is at least 49.9 EUR and the prod-
uct group code (WarennummerCode) is smaller than
62069090 (exluding some products for which return
patterns are presumably different) are predicted to be
returns. The antecedents of this rule cover 7,573 sam-
ples in training data of which only 1,469 (19.39%) are
not of class returns.
13
I.e. a 2.54% reduction in Area-under-ROC-curve
(AUC) versus the model trained on whole data (estimated
by two-fold CV. Precision and Recall cannot be directly
compared since these may be traded off differently in both
models.
14
All initial rules predict the non-default class, followed
by a rule with empty antecedents predicting the default
class.
(VKA.ARSeit_year >= 2014) and (VKA.VK_PreisA
>= 33.9) and (VKA.Laenge <= 0) and (VKA.
VK_PreisA >= 49.9) and (VKA.
EingebbarAb_weekday = Di) => class=1
(730.0/128.0)
Again, this rule only utilizes products base data
(VKA.∗). All products which have been listed since
year 2014 (ARSeit year) and have a sales price
(VK PreisA) of at least 49.9 EUR
15
and have length
(VKA.Laenge) of zero or less (excluding mainly
trousers since only these have positive values for
length) and have been set active in the webshop
(EingebbarAb weekday) on a Tuesday
16
are pre-
dicted to be returns. The antecedents of this rule cover
730 samples of which 128 (17.53%) are not returns.
(VKA.EingebbarAb_year >= 2016) and (VKA.
ARSeit_month >= 5) and (VKA.VK_PreisA >=
34.9) and (VKA.Laenge <= 0) => class=1
(2569.0/795.0)
Again this rule only utilizes products base data
(VKA.∗). All products that have been listed since
2016 and in each year listed from month May on-
wards (ARSeit month, including mostly those prod-
ucts sold in summer) and which have a sales price of
at least 34.9 EUR and a length of zero or less (again
excluding trousers) are predicted to be returns. The
antecedents of this rule cover 2,569 samples of which
795 (30.94%) are not returns.
(VKA.EingebbarAb_year >= 2016) and (VKA.
EingebbarAb_year >= 2017) and (VKA.
WarennummerCode <= 63012090) and (VKA.
WarennummerCode >= 61091000) => class=1
(368.0/120.0)
This is the last rule that only uses products base
data (VKA.∗). All products that have been listed
from 2017 and have a product group code
17
between
63012090 and 61091000 inclusive (this set consists
almost exclusively of blankets) are predicted to be re-
turns. The antecedents of this rule cover 368 samples
of which 120 (32.60%) are not returns.
(VKA.WarennummerCode >= 40169997) and (VKA.
ARSeit_year >= 2015) and (Auftrag.
Lieferdatum_weekday = Di) and (Auftrag.
LieferAdresseFl = 1) => class=1
(68.0/10.0)
This rule utilizes products base data (VKA.∗) as well
as order data (Auftrag.∗). All orders with products
15
Note that this feature appears twice in this rule but of
course just using the higher treshold value is equivalent to
the redundant form.
16
We presumed that different product groups were acti-
vated on different week days. However our project partner
could not confirm this.
17
This code is used for Intrastat declarations.
ICAART 2019 - 11th International Conference on Agents and Artificial Intelligence
806

having a product group code of at least 40169997
18
and have been listed since 2015 and have been de-
livered (Auftrag.Lieferdatum weekday) on a Tuesday
via express delivery (Auftrag.LieferadresseFl=1) are
returned. The antecedents of this rule cover 68 sam-
ples of which 10 (14.7%) are not returns. We hy-
pothesize that these orders were made over the week-
end and were expected to arrive on Monday but ar-
rived too late and were therefore returned. However
the small number of cases did not allow us to validate
this.
(VKA.WarennummerCode >= 40169997) and (Auftrag
.ErsterVersPlan_weekday = Do) and (Auftrag
.LieferAdresseFl = 1) => class=1
(107.0/13.0)
All orders containing products with a product group
code of at least 40169997 (see previous rule) and
which were last planned to be sent out on a Thurs-
day (Auftrag.ErsterVersPlan weekday) are predicted
to be returns. The antecedents of this rule cover 107
samples of which 13 (12.14%) are not returns. We
hypothesize this to be a variant of the previous rule.
(VKA.WarennummerCode >= 40169997) and (VKA.
EingebbarAb_year >= 2015) and (VKA.
ARSeit_year <= 2016) and (Auftrag.
Auftragsdatum_month <= 2) and (
Warenausgang.geliefert_weekday = Fr) =>
class=1 (74.0/20.0)
All orders containing products with a product group
code of at least 40169997 (see previous two rules)
and which have been set active from 2015 onwards
and have been available in the webshop until 2016
(i.e. they have only been available for 1-2 years)
and which were ordered in January or February (Auf-
trag.Auftragsdatum month <= 2) and were delivered
on a Friday (Warenausgang.geliefert weekday = Fr)
are predicted to be returns. The antecedents of this
rule cover 74 samples of which 20 (27.02%) are not
returns.
(VKA.WarennummerCode >= 40169997) and (Auftrag
.Lieferart = p) and (Auftrag.
ErsterVersPlan_weekday = Mo) and (Auftrag.
Kontaktform >= 12) => class=1 (144.0/33.0)
All orders containing products with a product group
code of at least 40169997 (see previous three
rules) and which have been delivered by post (Auf-
trag.Lieferart=p) and have last been intended to be
sent out on Monday and have been ordered in actual
shops (Auftrag.Kontaktform>=12, i.e. not ordered
via Webshop) are predicted to be returns. The an-
tecedents of this rule cover 144 samples of which 33
18
This excludes a small group of free product giveaways,
vouchers and made-to-order products and services which
are very unlikely or even impossible to return.
(22.91%) are not returns. Since this rule describes re-
turns outside of webshop orders it is not directly rele-
vant to our project.
(VKA.WarennummerCode >= 42023290) and (Auftrag
.ErsterVersPlan_weekday = Do) and (Auftrag
.Auftragsdatum_weekday = Di) and (Auftrag.
AuftragNichtTeilen = 0) => class=1
(94.0/14.0)
All orders containing products with a product group
code of at least 42023290 (this seems to basically ex-
clude similar products as the previous three rules with
this slightly different threshold) and which have been
planned to be sent out on a Thursday and which were
ordered on a Tuesday and which did not have the or-
der flag do not split (Auftrag.AuftragNichtTeilen=0)
are predicted to be returns. The antecedents of this
rule cover 94 samples of which 14 (14.89%) are not
returns.
(VKA.WarennummerCode >= 40169997) and (Auftrag
.Lieferart = p) and (Auftrag.
ErsterVersPlan_weekday = Mo) and (Kunden.
DatumErstanlage_year >= 2011) and (Auftrag
.Landkuerzel = a) and (Auftrag.
AufnahmeZeit_sec <= 23) => class=1
(117.0/33.0)
This is the first rule that also includes anonymized
customers base data (Kunden.∗). All orders con-
taining products with a product group code of at
least 40169997 (see some previous rules) and which
have been delivered by post (Auftrag.Lieferart=p) and
which have been last planned to be sent out on a Mon-
day, ordered by customers which were initially cre-
ated in 2011 or later, sent to an address in Austria
(Auftrag.Landkuerzel=a) and which were confirmed
in at most 23 seconds (Auftrag.AufnahmeZeit sec
<= 23) are predicted to be returns. The antecedents
of this rule cover 117 samples of which 33 (28.20%)
are not returns.
=> class=0 (10255.0/2538.0)
All samples not covered by any previous rule are pre-
dicted to be non-returns. This empty antecedent cov-
ers 10,255 samples of which 2,538 (24.74%) are re-
turns.
We note that the most features of these rules were
drawn from products base data (VKA.∗) followed by
orders (Auftrag.∗) and only the last (non-default-)rule
contains customers base data (Kunden.∗). This may
indicate the relative importance of these feature sub-
sets. We showed and explained these rules to our
project partner and they were quite surprised and in-
trigued by these results.
Towards Minimizing e-Commerce Returns for Clothing
807

6 CONCLUSIONS
We have discussed several approaches to reduce re-
turns in the context of garment e-commerce. A few
promising approaches could not be followed up be-
cause they proved too costly to implement, such as
adding garment size tags only after delivery and sub-
sequent measurement to effectively remove manufac-
turing tolerances. One simple approach – providing
more detailed product-specific measurement tables –
is currently in evaluation. We shortly mentioned the
usefulness of qualitative size information and pre-
sented some preliminary results.
In the main part of our paper, we describe a
method to identify and remove highly predictive fea-
tures from large, mostly undocumented datasets to
improve the quality and stability of trained models
while also preventing overfitting. We demonstrate the
usefulness of this method by describing a rule set of
only eleven rules that predicts returns at good preci-
sion and recall on a large real-life dataset. To achieve
this, it was also necessary to modify the chosen learn-
ing algorithm RIPPER in a minor way to ensure it
always characterizes returns rather than non-returns.
The described rules show some intriguing patterns
which are currently investigated by our commercial
partner and some may prove to be generally useful.
In the future we hope to follow up on reconstruct-
ing precise body size from ordering information – ob-
serving that we already obtained reasonably precise
manufacturing size information – and finish our pre-
liminary investigations towards a final result.
ACKNOWLEDGEMENTS
This project was funded by the Austrian Research
Promotion Agency (FFG) and by the Austrian Fed-
eral Ministry for Transport, Innovation and Technol-
ogy (BMVIT) as project Think!First (859099)
REFERENCES
Cohen, W. (1995). Fast effective rule induction. In Pro-
ceedings of the Twelfth International Conference on
Machine Learning, San Francisco, CA, 1995, pages
115–123. Morgan Kaufmann.
Colsen, E. (2013). Using human and machine processing
in recommendation systems. In Human Computation
and Crowdsourcing: Works in Progress and Demon-
stration Abstracts, number 13-01 in CR.
Ghaffari, S. (2011). Will it fit? consumer decision mak-
ing in online shopping environments. Master’s the-
sis, School of Industrial Design, Georgia Institute for
Technology, USA.
Hagemann, H. (2015). Umweltrelevante Produktinfor-
mationen im E-Commerce: Chancen f
¨
ur nachhalti-
gen Konsum. Umweltbundesamt, D-06844 Dessau-
Roßlau, Germany.
Halbach, J., St
¨
uber, E., and Piepke, M. (2015). Nach-
haltigkeit im Online-Handel - Die Rolle von Aus-
gestaltung und Kommunikation. IFH Institut f
¨
u Han-
delsforschung GmbH.
Hofacker, L. and Langenberg, C. (2015). E-Commerce-
Markt
¨
Osterreich/Schweiz 2015. EHI Retail In-
stitute. https://www.ehi.org/deStudien/e-commerce-
markt-oesterreichschweiz-2015.
Knabl, W., K
¨
ob, M., Meszaros, G., Prenger, C., Rischanek,
U., Segal, D., and Weigl, A. (2015). retail - Magazin
f
¨
ur den
¨
osterreichischen Handel. Offizielles Medium
des Handelsverbandes, volume 04. Handelsverband.
https://www.handelsverband.at/fileadmin/content/
images publikationen/retail/Retail 2015 04.pdf.
Kristensen, K., Borum, N., Christensen, L., et al. (2013).
Towards a next generation universally accessible
online shopping-for-apparel system. In Human-
Computer Interaction: Users and Cotexts of Use, Vol-
ume 8006 of the series Lecture Notes in Computer Sci-
ence, pp. 418-427. Springer.
Lengauer, E., Koll, O., Kreuzer, M., Herry, M., and Sed-
lacek, N. (2015). eComTraf - Auswirkungen von E-
Commerce auf das Gesamtverkehrssystem. Austrian
Federal Ministry for Transport, Innovation and Tech-
nology (BMVIT). Study funded within the search pro-
gram Mobility for the Future (Mobilit
¨
at der Zukunft,
MdZ) by the Austrian Federal Ministry for Transport,
Innovation and Technology (BMVIT).
Seewald, A. (2007). Improving the effectiveness of mail-
ings by building a response model for inactive cus-
tomers. Technical Report 2007-01, Seewald So-
lutions, L
¨
archenstraße 1, A-4616 Weißkirchen a.d.
Traun, Austria.
Singh, K. (2015). Reducing Customer Returns in an Online
Marketplace. Dept. of Fashion Technology, National
Institute of Fashion Technology, Mumbai, India.
Toktay, L. (2003). Forecasting product returns. In Guide Jr.,
D. and Van Wassenhove, L., editor, Business Apects
of Closed-Loop Supply Chains, International Man-
agement Series, volume 2. Carnegie Bosch Institute.
Wernbacher, T., Pfeiffer, A., Denk, N., Platzer, M., Berger,
M., Seewald, A., Winter, T., and Miller, I. (2017).
Minimizing returns through gamification, persuasive
design principles & machine learning. In Poster pre-
sentation at the 11th European Conference on Games
Based Learning (ECGBL 2017), Oct. 5-6 2017, Graz,
Austria.
Willemsen, R., Abraham, J., and van Welle, R. (2016).
Global B2C E-commerce Report 2016. Ecommerce
Foundation, Raddhuisstraat 22, NL-1060 Amsterdam,
The Netherlands.
ICAART 2019 - 11th International Conference on Agents and Artificial Intelligence
808