https://doi.org/10.31449/inf.v45i4.3066 Informatica 45 (2021) 625–632 625
A Practical Framework for Real Life Webshop Sales Promotion Targeting
Gábor K˝ orösi and Tamás Vinkó
University of Szeged, Institute of Informatics, Hungary
E-mail: korosig@inf.u-szeged.hu,tvinko@inf.u-szeged.hu
Technical paper
Keywords: promotion targeting, behavior analysis, hybrid recommendation system
Received: February 28, 2020
In recent years, online marketing has become increasingly extensive and effective. Product recommender
systems are often deployed by e-commerce websites to improve user experience and increase sales. To
address this, more and more e-commerce started to use machine learning models to predict customers’
purchase behaviors. In the scientific literature there are only few real-life studies to date which give solu-
tions for recommendation systems for online advertising. The demand from the owners of such websites
is given, however, it is hard for them to choose a method or model to predict from an endless number of
options for some specific circumstances. The aim of this paper is to propose a practical guideline as a
hybrid approach that predicts customers’ purchase behaviors and helps to target advertisement, sales form
in user level. To this end, we have designed a robust hybrid model to predict interested sales form based on
user behavior within a large e-commerce website. The paper details a real-life practical solution and build
a structure that can be used in a large variety of e-commerce systems.
Povzetek: Opisan je razvoj modela nakupovanja po spletu z namenom ciljnega oglaševanja.
1 Introduction
One of the most important and dynamically developing ar-
eas today is e-commerce and related services. While in a
traditional shop tracking customer is difficult (e.g.,loyalty
card program), a webshop’s back-end offers countless so-
lutions to solve this problem. For example, we could use
cookies, spent checking, newsletter and product tracking
[4, 15, 1]. The main driving force behind this fast evolu-
tion is the fact that we can understand and anticipate user
behavior better, and we can answer the related questions in
real-time. The key goal is to get the highest response from
users by spending as little money and time on it as possible,
and create customer-oriented services [2]. That is named
personalization and targeting [13], where the objective is
to find the best matching ads or form of sales promotion to
be displayed for each user. The solution is not new, as we
could see similar solutions at the first generation webshops,
but nowadays the amount of data is much higher than be-
fore.
When the task is to efficiently process huge amount of
data, it is useful to try and find a solution in those research
papers written based on similar task. For example, [26] an-
alyzing clickstream, [15] uses email sending history, while
[1] collects user activity to predict user’s future behavior.
At first glance, the task does not seem to be a difficult one,
as using data mining and data science in e-commerce is not
new, and there is a huge amount of papers published with
the same purpose. These papers refer to a waste amount of
machine learning (ML) tools and solutions which are able
to help with this optimization. For instance, classification
can predict the occurrence of an event, or regression tech-
niques that can help us to predict the time or amount of
money the user will spend on the website. More sophis-
ticated solutions are offered by collaborative filtering or
content-based approaches. The repository of toolkits may
seem endless, but solving a problem is never the same, and
it is seldom enough to use just one tool to solve a problem.
Many recent publications have introduced some kind of hy-
brid solution for this complex problem in which one has to
combine and embed simple methods to find a proper model.
For example, in [5] we could see a typical hybrid recom-
mendation model that integrates user-based and item-based
collaborative filtering, content-based filtering together with
contextual information to get rid of the disadvantages of
each approach.
Thorough literature review on these subjects can lead to
an impression that most of the scientific papers are theo-
retical model descriptions instead of accurate and practical
model descriptions. One could find vague model formula-
tions that make it difficult or impossible to rebuild a pre-
sented solution in real life. Along this line of thought, we
have concluded that, besides theoretical models, there is a
huge demand for publications that document a case study
and provide the opportunity for anyone to reproduce it the
results on their database. Our goal is to make and docu-
ment a case study that demonstrates an ML-based recom-
mendation system, which classifies users and provides an
individual-level approach for ads form. Based on our litera-
ture review we found that a hybrid recommendation system

626 Informatica 45 (2021) 625–632 G. K˝ orösi et al.
provides the most accurate solution for that. We combined
and embedded various classification and regression mod-
els, including Logistic Regression, Random Forest, GBM,
and XGBoost to get the most accurate solution.
The rest of the paper describes our approach as follows.
Sections 2 and 3 describe the background of the problem.
In Section 4 the dataset and the generated features are de-
tailed. The model ensemble is briefly described in Section
5. In Section 7, the importance of features is studied, top
features are listed, and our solution is given. Finally, Sec-
tion 8 concludes the results of the study.
2 Background
As data is increasing, more and more companies are de-
manding high quality solutions from their data scientists.
The use of recommendation systems has become a daily
concept in product suggestion, product group selection,
promotional message content generation which is sup-
ported by machine learning techniques. Common exam-
ples of applications include the recommendation of movies
(e.g., Netflix, Amazon Prime Video), music (e.g., Pandora),
videos (e.g., YouTube), news content (e.g., Outbrain) or ad-
vertisement (e.g., Google) [25].
In this paper we give a detailed description of a rec-
ommendation system which can make user-level marketing
letter or offer sales promotion. Note that recommendation
system is a quite general concept. It could be based on the
collaborative filter solution, the content-based method, the
classification or regression, and their embedding in differ-
ent depths and widths. What follows is an outline of what
a recommendation system might consist of.
Collaborative filtering (CF) is probably one of the most
used and well-known technologies. Behind the basic idea,
the solution is that based on users’ historical data, the users
are put into an n-th dimensional space which makes pos-
sible to then measure the distance between them. In light
of this, we could make recommendations based on the data
of the users closest to each other [7]. This CF technique
proved its power, but on the other hand, a huge amount of
work pointed out the disadvantages of it. These are the fol-
lowings: cold start problem, data sparsity, and scalability
[29].
Besides collaborative filtering, the second most popu-
lar solution is the content-based method. It is a tech-
nique which operates with unique characteristics and be-
haviors of each customer, and in turn, delivering personal-
ized content for each user, based on their content consump-
tion history across channels. Another interesting way is the
community-based method. This approach assumes that the
content coming from a user’s friends or authoritative users
is more likely to be interesting for a user than the rest.
While collaborative filtering and content-based models,
used only a static ’user states’ we could find many papers
which are using uni- or multivariate user event sequences,
time-series to build a predictive model. Koehn et al. [20]
divided the user event sequence prediction problem into
four groups, namely the ’predict the product group’, ’clas-
sify a sequence’, ’predict the outcome of an incomplete ses-
sion’, and ’click-through rate prediction’. In our work we
are focusing to predicting the users’ interest, which was
created based on some initial observations on the users’
purchase behavior during the shopping process, meaning
that our task is rather similar to the ‘predict the product
group’ task of the recommender systems. Koehn et al. [20]
summarized the methods of event sequence data prepro-
cessing, highlighting their advantages and disadvantages.
One of the most often implemented methods is to create ag-
gregated, cumulated data, which, however, results in data
loss and requires manual feature engineering by the do-
main experts. Another common method is to create se-
quence segments or sliding a window, where we use only a
chunk/fixed-length part of the data. Lastly, there are neural
networks and embedding layers, where we can work with
partially or completely raw data. In the field of sequence
prediction approach, we could find many papers.
Perhaps one of the most promising paper which related
to our work is created by Yu et al. [28]. They used re-
current neural network on sequenced data to identify web
shop users habits and made the next basket recommenda-
tion. They applied recurrent layers in the temporal do-
main and proved their effectiveness for handling the tem-
poral dimension for time series classification. Deep learn-
ing based (DLL) solution with time series have proven ef-
ficiency in many areas, however, web-shop log data often
includes variables that contain mixed continuous and dis-
crete variables. Even these kinds of data can be easily han-
dled by a decision tree-based solution, in neural network
this is not so easy. In deep learning based approaches, the
discrete-valued sequences must be transformed into the nu-
meric space. Using one-hot encoding might not prove to
be overly useful, as it explores the dimensionality of the
input feature vector and dramatically increases its sparsity.
Inspired by Natural Language Processing, we managed to
transform our categorical data into a dense space utilizing
embeddings. These methods encode categories as vectors
based on contextual similarities and then feed them into
the recurrent or convolutional neural network. The em-
bedded vectors are usually trained together with the time-
series/sequence model training process [21]. The embed-
ding of discrete-valued sequences was successfully applied
in user behavior analysis. For instance, An et al. [3] pre-
sented their neural user embedding approach which was
capable of learning informative user embeddings by us-
ing the unlabeled browsing-behavior. Koehn et al. [20]
proposed their impressive clickstream classification results
where they applied RNN architectures and embedding lay-
ers. Cheng et al. [9] introduced the Wide and Deep fea-
ture representation method. In their terminology, Wide rep-
resentations were one-hot encoded features which could
memorize sparse feature coincidences, while Deep repre-
sentations consisted of dense embeddings which gave gen-
eralization power to deep learning systems.

A Practical Framework for Real Life Webshop Sales. . . Informatica 45 (2021) 625–632 627
Although content-based and community-based methods
have proven their worth in many case studies, in our case
it was almost impossible to apply these methods due to the
lack of data. Another approach could be the deep learning
based solution, but as many paper shows (e.g. [8, 17]) when
the dataset is based only short sequences (as our dataset), a
traditional ML model can outperform a DLL based model.
Based on these paper even a XGboost based classifica-
tion model or regression would provide a good solution in
an optimum prediction system, but unlike simple patterns,
things are always more complicated in real life.
To solve the backward of the traditional and DLL based
methods, the concept of hybrid or combined systems are
becoming more popular in many papers. Bozanta and
Kutlu [5] summarized that while each filtering approach
has different drawbacks, a hybrid approaches combines the
existing approaches and aim to minimize or remove the
drawbacks of existing approaches, which may occur when
they are used individually. The exact description does not
exist for a hybrid solution, but we could certainly use the
aforementioned tools at different depths and widths. There
are quite many papers proving that a hybrid approach pro-
vides a better solution than the single method, see, e.g., in
[5, 7, 12]. Thus, we have chosen this solution for our work-
flow, and we decided to use use such a hybrid model for
our system which used both regression and classification
method.
Our goal was to solve the problem of predicting the
user behaviors about the sales promotion. Similar goals
is solved by Martínez et al. [23] and Liu et al. [22]. They
created a model that can predict future customer behavior
which based on the set of customer-relevant features that
derives from times and values of previous purchases. As
our solution, they apply machine learning algorithms in-
cluding logistic Lasso regression, the extreme learning ma-
chine and gradient tree boosting for predicting whether the
customer makes a purchase in the upcoming month. Al-
though these two cited papers are very similar to the so-
lution we used, however, unlike them, we tried to create a
prediction algorithm not just by using one method but by a
(hybrid) combination of them.
3 Problem statement
The main objective of this paper is to solve the problem of
predicting the purchase behaviors of users who have known
the history on an e-commerce website. More closely, we
aim at forecasting which ads group or form of sales pro-
motion user will most likely to use based on purchase his-
tory and profile information. This form of sales promotion
could be: buy two, get one free; price deal; sampling, etc.
Although we did not directly use others’ work to design
our system, the solution we came up with is strikingly sim-
ilar to the description of [29]. That is, a predictive system
would help in several practical scenarios such as
– build a cold start recommender system, by providing
high-level recommendations to users who connect for
the first time to an e-commerce website;
– improve existing product recommendation engines,
by providing category-level priors that can guide the
recommender system to and domains of interest for
the user;
– provide e-commerce companies with tools for tar-
geted email/social media campaigns.
Our paper has two main goals. The first is to explore
which information is correlated with the form of sales pro-
motion which the users most likely to use (see in Table 1
for an illustrative example.) Based on this we have built
and tested a hybrid model which optimizes a user-level ta-
ble, in order to propose the form of sales promotion to users
that fit the best to their interests and preferences, see Table
2. The second goal is to back-test and document well each
critical point of hybrid machine learning algorithms which
could be used as a base structure for those who want to
replicate our model or build a similar system.
4 Datasets
We have used data which has been recorded from a health
and beauty webshop. The data has contained near millions
of users, from different markets (countries), however, in
order to obtain the richest data possible, we have filtered it
by the oldest market which includes 230; 000 user-profiles
and their purchase history. Data consists of seven years of
user interaction logs with the webshop. Each event has a
user identifier, a timestamp, and an event type. The pur-
chase data contains 5 categories of events: pageview of a
product, basket view, buy, ordered timestamp, and deliv-
ered timestamp.
There are around 240 different types of products. In the
case of a buy or a basket view, we have information about
the price and extra details. An average customer has been
used the shop two or three times yearly, which leads to very
sparse and high dimensional dataset. This is not surprising
as it is extremely common in recommender systems [25].
As a solution, there are two obvious ways to reduce the di-
mensionality of the data: either by marginalizing the time
(aggregate pageviews per user over the period) or the prod-
uct pageviews (aggregate products viewed per time frame)
[26]. In this work, we follow both approaches.
As a first step, our solution connected unique events with
sessions. We used homogeneous like purchase history only
and heterogeneous example clicks, profile data in nature.
These events are then cleansed and ordered by their times-
tamps to form the action chain.
As a next step, we transformed unique events into a fea-
ture list (e.g., number of purchases, the distance between
two logins, etc.). Beside of the evident data (number of,
sum of, mean of purchases), the script accumulated other
data such as:

628 Informatica 45 (2021) 625–632 G. K˝ orösi et al.
Table 1: Illustration: problem statement as a binary classification.
1st 2nd 3rd nth
purchase purchase purchase  purchase
Likely to buy with
time of sales promotion
prediction ) (user who use more
than 50% promotion
for buying something)
Table 2: Illustration: problem statement as a recursion; the distribution of sales promotion types.
1st 2nd 3rd nth
purchase purchase purchase  purchase
SPType1 35%
Time of ) SPType2 25%
prediction
.
.
.
.
.
.
SPTypen 50%
– distance (in time) between first and second, third,
etc. actions;
– number of purchases in first, second, etc. months;
– increase or decrease in purchases compared to the pre-
vious month by month;
– the reaction times between advertising letters and a
purchase.
4.1 Feature engineering
One of the most important steps for better performance
of a classifier is to preprocess the data correctly. Besides
the regular data cleaning process, we transformed features
by scaling each feature to a given range with min-max
scaling. As a last preprocessing step, we calculated fea-
ture importance with tree-based ensemble method namely
ExtraTreesClassifier [14]. Based on the obtained
results by this method, our model uses only the top 20 fea-
tures, which significantly increased the accuracy of the re-
sults.
5 Methodology
In order to handle the popularity-bias, we divided the prob-
lem into two subtasks:
i) predict if a user is sensitive for the sales promotion or
not, and
ii) predict which kind of form of sales promotion is more
interested in it.
As a solution, we have created a hybrid model which used
both regression and classification method, see Figure 1.
The recommendation model returns two lists. The first
list gives information about the users, if they are likely to
use or not any of the forms of sales (the sensitivity for sales
promotion). The second list provides us with the data to
calculate the probability for every sale (which form of sales
likely to use).
For the results we propose a novel hybrid recommenda-
tion algorithm where similarity measurement is performed
between a user and form of sales on features derived from
their profile and history information. As a result, we obtain
a table where every single user gets his/her predicted value,
as we can see in Table 3.
6 Experimental setup
As we want to use raw log data to make a prediction for
recommendation, we have to handle the data sparsity prob-
lem. As already mentioned, our dataset contains 230; 000.
However, only 33; 000 of them have data of sufficient qual-
ity. So, in our experiment, we used only this reduced and
filtered dataset. To conduct experiments, we split the entire
dataset into test (20%) and training (80%) sets.
In the first step, we have trained various classification
models, including Logistic Regression [11], Random For-
est [6], LightGBM [19], and XGBoost [8], where grid
search was used to select the optimal parameters. As the fi-
nal results proved, XGBoost classifier and XGBRegressor
performed the best. Additionally, the majority of classifier
(MC) [18] is used as a baseline for comparison with the
above learning algorithms.
For the regression problem, we used the central ten-
dency measure as the baseline for all predictions. Based
on these, we inspected the hybrid models using the train-
ing set and adjusted the predictive algorithms’ parameters
achieving the best performance on the validation set. Pre-
dictions were made for each instance in the test set and the
forecasted results were compared with the true values by
computing corresponding performance metrics. To obtain
the best evaluations we have usedK-fold validation where

A Practical Framework for Real Life Webshop Sales. . . Informatica 45 (2021) 625–632 629
Figure 1: State diagram of our hybrid solution.
Table 3: Example of model outcome.
user id likely to use likely to use sales promotion type
sales promotions type1 type2 type3 type4 type5
1000 YES 35% 50% 5% 3% 7%
1001 NO 0% 0% 0% 0% 0%
both training and validation sets were also used for predic-
tion.
Handling problem with an ensemble
classification and regression tree
The first goal is to predict if a user is likely to use or not a
sales promotion, which is a binary classification problem.
To find the best solution we have trained and tested classi-
fication models as many times as we could. In the end, we
have found that the XGBoost ensemble classifier [8] gives
the best results. It is not surprising, because tree boost-
ing is a highly effective and widely used machine learning
method.
Another important feature is that the algorithm has a
good performance as it includes an efficient linear model
solver and can also exploit parallel computing capabilities
[8]. Ensemble learning to provide a systematic solution to
merge the power of multiple learners. The prediction value
of XGB can have different interpretations, depending on
the task, i.e., regression or classification. XGB is a tree
ensemble model which set of classification and regression
trees. It could classify our data into one of a finite number
of values, that while called a regression (nonlinear model).
Besides XGB, we compared our results with Linear regres-
sion [10], Lasso [24] and Ridge regression [16].
7 Results
Classification. It is well known that the main problem of
the recommendation system is the cold start problem. It
could appear when the user has started his/her initial steps,
or in our case when a shop owner started a new sales pro-
motion type, which makes very sparse data. To solve this
problem, we filtered (dropped out) those users and promo-
tions from the training dataset which has too sparse or no
data. Based on our model, we made a binary classification
with XGBoost to predict user likely to use a sales promo-
tion or not. The parameters of the estimator used to apply
optimization by cross-validated grid-search over a parame-
ter grid.
To find the most accurate model, we have tried more
models and settings. The results are reported in Table 4,
where the window size (number of purchase) was 3 for all

630 Informatica 45 (2021) 625–632 G. K˝ orösi et al.
Table 4: Results of classifications.
model ACC F1 precision recall
Baseline 0.587 0.342 0.351 0.337
Logreg_all 0.676 0.409 0.620 0.306
Logreg_top10 0.685 0.404 0.661 0.291
XGBoost_all 0.706 0.527 0.652 0.436
XGBoost_top10 0.703 0.518 0.657 0.419
XGBoost_all_HPT 0.768 0.519 0.666 0.423
XGBoost_top10_HPT 0.771 0.509 0.658 0.417
XGBoost_top10_HPT(4) 0.790 0.624 0.713 0.554
Table 5: Error rates of regression models.
model Sales promotion Type1 Sales promotion Type2 Sales promotion Type3
MAE MSE RMSE MAE MSE RMSE MAE MSE RMSE
Baseline_CV 5.840 53.568 7.313 9.970 161.948 12.723 12.679 256.261 16.001
DNN 5.906 53.275 7.298 9.870 158.492 12.589 12.600 259.132 16.097
LR_all_CV 5.039 46.029 6.779 8.927 131.045 11.442 11.368 206.408 14.365
LGBMReg_CV 4.715 42.469 6.551 8.446 118.202 10.869 10.946 191.677 13.843
StackReg_CV_TOP 4.778 43.153 6.564 8.720 125.506 11.200 11.092 196.392 14.013
LR_CV_TOP 4.986 44.844 6.691 8.829 127.842 11.301 11.234 203.471 14.284
LGBMReg_CV_TOP 4.700 42.349 6.501 8.602 112.589 11.067 10.895 191.120 13.824
the methods, except in the last configuration.
During the first phase, we have used XGboost with all,
and with only the top-10 features, which achieved 70% ac-
curacy. To improve this, we have applied hyperparameter
tuning, namely cross-validated grid-search over a parame-
ter grid which gains better accuracy.
We wanted to make further improvements, but the spar-
sity of the data did not allow it. The main problem is that
we want to predict user feature habits as soon as possible.
For that reason, we used the user’s first 3 purchase history
to train the model, but this was (as expected) not enough
to improve the results. To get better results, we need more
data, such as we expected it. The solution to this problem
is simple: we have to wait for more information, or encour-
age clients to fill the profile table. To prove this concept,
we have trained our model with the user’s first 4 purchases,
which achieves 0:79 accuracy (last row in Table 4).
To find another way for this challenge, we changed our
method like many researchers suggest: if we don’t have ac-
curate enough classification model, we have to change our
point of view. To use this idea, we retested our solution as a
regression with XGBRegression (as a regression problem).
As a result, it affordsRMSE = 16:77, which is not offer-
ing better outcome, because if we transform this result into
a classification result, we got accuracy: 0:686, precision:
0:578, recall: 0:546, and F1: 0:562.
Regression. In our second phase, we were looking for-
ward to determining which type of sales promotion will
prefer most of our users (see in Figure 1). It is a regres-
sion problem, where we have to predict every SP type for
every user. To make a measurable result, we did not test all
the types of SP, instead of that, we chose only 3 types of
promotion:
– Type1 is an SP type which has a long history in our
webshop;
– Type2, which has only a year background, and
– Type3 is the youngest SP type (less than 6 months is
using).
Based on this idea we obtained the results reported in Table
5.
To get the best outcome, we tested more models with
different settings, like linear regression (LR), LightGBM
(LGBMReg), and a simple deep neural network (DNN).
In the initial step, our model used all (n = 129) normal-
ized, scaled and skewed feature sets. Based on this method
LGMBReg made the most accurate solution.
As a second step, we wanted to increase our model’s ac-
curacy. To solve this, we wanted to find the most impor-
tant features. For this purpose, we used wrapper method,
namely backward elimination. As the name suggests, we
gave all the possible data to the model at first. We track
the performance of the model and then repetitively remove
the worst performing features one by one until the over-
all performance of the model comes in a suitable range.
To calculate feature importance, we are using the ordinary
least squares (OLS) model [27]. After many attempts and
settings, the best solution is made by LGMBReg which is
a tree-based regression model, which made a much more
accurate model than the random choice.

A Practical Framework for Real Life Webshop Sales. . . Informatica 45 (2021) 625–632 631
Discussion. Our problem and its solution to predict ac-
ceptance of the sales promotion is unique, since we do not
predict a repeat purchase but a reaction to advertising let-
ters. Regardless, we wanted to somehow compare the per-
formance of our model with other models as well. The re-
sults, and methodology of our paper is similar to the results
obtained by Martínez et al. [23], so we compared our re-
sults with theirs. Our goal was to predict if a user is likely
to use or not a sales promotion, which was same as their bi-
nary classification problem. While our model reaches 79%
accuracy, their solution reached 86.68%. The difference in
accuracy between the two models is not surprising, since
we used only the first 4 purchases, they used 24 months for
the same task. As they noted, it is difficult to make an ac-
curate prediction model from short data and few purchases,
however, over time, as data is collected, we could produce
more accurate results.
8 Conclusions
In this work, the goal was to build and share a structure of
the model for predicting user habits about using sales pro-
motions. As we saw in the literature review it is not a trivial
case. There is a lot of gaps that we have to handle, for ex-
ample, feature with different types (time, numeric, categor-
ical, etc.) or scale. Based on human habits, the webshop’s
data is often log scaled, and sparse which makes it difficult
for the model to find optimal parameters. There are now
countless solutions to deal with this problem, like scaling,
normalizing, skewing data, or find the most relevant fea-
tures. Based on these methods, finally we identified a solu-
tion for our problem with relatively good accuracy results.
For the classification problem we have found that XGBoost
gives the best model, while the second solution is not that
clear. Based on our results as at first glance LightGBM
(LGBM) could be the right choice.
Before making our decision, we need to know the struc-
ture of the model. LGBM is a very popular solution, be-
cause of its speed and accuracy. It has happened because
LGMB grows tree vertically while other algorithms like
Xgboost, Gboost grow trees horizontally. Put it differently,
LGBM grows tree leaf-wise while another algorithm grows
level-wise. LGBM is giving the best solution for our task,
but there is some gap, which overshadows our success.
However, it is sensitive to overfitting, especially on small
dataset. There is no threshold on the number of rows but
researchers suggest to use it only for data with 10; 000+
rows. This model hence cannot be used for new promo-
tions that only used by a small amount of user. In the light
of this, in the final model, we used linear regression which
gives almost the same results as LGBM.
References
[1] Ahmed, A., Low, Y ., Aly, M., Josifovski, V ., and
Smola, A. J. Scalable distributed inference of dy-
namic user interests for behavioral targeting, In:
Proceedings of the 17th ACM SIGKDD international
conference on Knowledge discovery and data mining,
2011
https://doi.org/10.1145/2020408.2020433
[2] Aly, M., Hatch, A., Josifovski, V ., and
Narayanan, V . K. Web-scale user modeling
for targeting. Proceedings of the 21st Interna-
tional Conference on World Wide Web, 2012
https://doi.org/10.1145/2187980.2187982
[3] An, M., Kim, S. Neural User Embedding
from Browsing Events. In: Machine Learn-
ing and Knowledge Discovery in Databases: Ap-
plied Data Science Track. ECML PKDD 2020
https://doi.org/10.1007/978-3-030-67667-4_11
[4] Banerjee, A., and Ghosh, J. Clickstream Clustering
using Weighted Longest Common Subsequences. In:
The Web Mining Workshop at the 1st SIAM Confer-
ence on Data Mining, 2001
[5] Bozanta, A., and Kutlu, B. Developing a Contextu-
ally Personalized Hybrid Recommender System, Mo-
bile Information Systems, Article ID 3258916, 2018
https://doi.org/10.1155/2018/3258916
[6] Breiman L. Random forests - random features. Tech-
nical Report 567, Statistics Department, University of
California, Berkeley, 1999
[7] Burke, R. Hybrid recommender systems: Sur-
vey and experiments. User modeling and
user-adapted interaction, 12(4), 331-370, 2002
https://doi.org/10.1023/A:1021240730564
[8] Chen T. and Guestrin C. XGBoost: A Scalable
Tree Boosting System. Proceedings of the 22nd
ACM SIGKDD International Conference on Knowl-
edge Discovery and Data Mining, pp. 785-794, 2016
https://doi.org/10.1145/2939672.2939785
[9] Cheng, H., Koc, L., Harmsen, J., Shaked, T., Chan-
dra, T., Aradhye, T., Anderson, G., Corrado, G., Chai,
W., Ispir, M., Anil, R., Haque, Z., Hong, L., Jain, V .,
Liu, X., Shah H. Wide & Deep Learning for Recom-
mender Systems. In: Proceedings of the 1st Work-
shop on Deep Learning for Recommender Systems,
2016 https://doi.org/10.1145/2988450.2988454
[10] Cook R.D. Detection of influential observations in
linear regression. Technometrics, 19(1):15-18, 1977
https://doi.org/10.2307/1268249
[11] John N. D. and Ratcliff D. Generalized iterative scal-
ing for log-linear models. The Annals of Mathemati-
cal Statistics, 43(5):1470-1480, 1972

632 Informatica 45 (2021) 625–632 G. K˝ orösi et al.
[12] Çano, E., Morisio, M. Hybrid recommender
systems: A systematic literature review. In-
telligent Data Analysis, 21(6), 1487-1524, 2017
https://doi.org/10.3233/IDA-163209
[13] Essex, D. Matchmaker, matchmaker. Com-
munications of the ACM, 52(5):16-17, 2009.
https://doi.org/10.1145/1506409.1506415
[14] P. Geurts, D. Ernst., and L. Wehenkel, Extremely ran-
domized trees. Machine Learning, 63(1), 3-42, 2006.
https://doi.org/10.1007/s10994-006-6226-1
[15] Grbovic, M., Radosavljevic, V ., Djuric, N., Bhamidi-
pati, N., Savla, J., Bhagwan, V ., and Sharp,
D. E-commerce in Your Inbox: Product Rec-
ommendations at Scale. In: Proceedings of
the 21th ACM SIGKDD International Conference
on Knowledge Discovery and Data Mining, 2015
https://doi.org/10.1145/2783258.2788627
[16] Hoerl, A. E., Kennard, R. W., Ridge Re-
gression: Applications to Non-Orthogonal
Problems. Technometrics 12(1), 69-82, 1970
https://doi.org/10.2307/1267352
[17] A. Ibrahem, A. Osman, A. N. Ahmed, M. F. Chow,
Y . F. Huang, A. El-Shafie. Extreme gradient boosting
(Xgboost) model to predict the groundwater levels in
Selangor Malaysia, Ain Shams Engineering Journal,
12(2), 1545–1556, 2021
[18] James, G. Majority vote classifiers: theory and appli-
cations. PhD thesis, Stanford University, 1998
[19] Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma,
W., Ye, Q., and Liu, T. Y . LightGBM: A Highly Effi-
cient Gradient Boosting Decision Tree. In: Advances
in neural information processing systems, pp. 3146-
3154, 2017.
[20] Koehn, D., Lessmann, S., Schaal, M. Pre-
dicting online shopping behaviour from click-
stream data using deep learning. Expert Sys-
tems with Applications, 150, 113342, 2020
https://doi.org/10.1016/j.eswa.2020.113342
[21] Li, Z., Kulhanek, R., Wang, S., Zhao, Y ., Wu, S.
Slim Embedding Layers for Recurrent Neural Lan-
guage Models. In: Thirty-Second AAAI Conference
on Artificial Intelligence. 2018.
[22] G. Liu, T. T. Nguyen, G. Zhao, W. Zha, J. Yang,
J. Cao, M. Wu, P. Zhao, W. Chen. Repeat Buyer Pre-
diction for E-Commerce. In Proceedings of the 22nd
ACM SIGKDD International Conference on Knowl-
edge Discovery and Data Mining, KDD ’16, 2016.
https://doi.org/10.1145/2939672.2939674
[23] A. Martínez, C. Schmuck, S. Pereverzyev,
C. Pirker, M. Haltmeier. A machine learning
framework for customer purchase prediction in
the non-contractual setting. European Journal
of Operational Research, 281(3)588–596, 2020
https://doi.org/10.1016/j.ejor.2018.04.034
[24] Park, T., Casella, G., The Bayesian
Lasso. Journal of the American Statisti-
cal Association 103(482), 681-686, 2008
https://doi.org/10.1198/016214508000000337
[25] Sidana, S. Recommendation systems for online ad-
vertising. Computers and Society [cs.CY]. Université
Grenoble Alpes, 2018.
[26] Vieira, A. Predicting online user behaviour us-
ing deep learning algorithms. arXiv preprint
arXiv:1511.06247, 2015
[27] Weiss, A. A Comparison of Ordinary Least
Squares and Least Absolute Error Estima-
tion. Econometric Theory, 4(3), 517-527, 1988
https://doi.org/10.1017/S0266466600013438
[28] F. Yu, Q. Liu, S. Wu, L. Wang, and T. Tan. A dy-
namic recurrent model for next basket recommen-
dation. In: Proceedings of the 39th International
ACM SIGIR Conference on Research and Develop-
ment in Information Retrieval, pp. 729–732, 2016
https://doi.org/10.1145/2911451.2914683
[29] Zhang, Y ., and Pennacchiotti, M. Predicting purchase
behaviors from social media, In: Proceedings of the
22nd International Conference on World Wide Web,
2013 https://doi.org/10.1145/2488388.2488521