Predicting Shopping Intent of e-Commerce Users using LSTM

Recurrent Neural Networks

Konstantinos Diamantaras

, Michail Salampasis

, Alkiviadis Katsalis

and Konstantinos Christantonis

Intelligent Systems Laboratory, Department of Information and Electronic Engineering,

International Hellenic University, Sindos, Thessaloniki, Greece

Keywords: Purchase Intent, e-Commerce, LSTM-RNN, Web Usage Mining.

Abstract: An e-commerce web site is effective if it turns visitors into buyers achieving a high conversion rate. To this

realm, it is useful to predict each user’s purchase intent and understand their navigation behavior. Such

predictions may be utilized to improve web design and to personalize shopper’s experience, hopefully leading

to increased conversion rates. Additionally, if such predictions can be done in real-time, during the ongoing

navigation of an e-commerce user, the e-commerce application can take proactive stimuli actions to offer

incentives with a view to increase the probability that a user will finally make a purchase. This paper presents

a method for predicting in real-time the shopping intent of e-commerce users using LSTM recurrent neural

networks. We test several variants of our method in a dataset created from the processing of Web server logs

of an industry e-commerce web application, dividing user sessions in three different classes: browsing, cart

abandonment, purchase. The best classifier achieves a predictive accuracy of almost 98%. This result is

competitive with other state-of-the-art methods, which affirms that accurate and scalable purchasing intention

prediction for e-commerce, using only session-based data, is feasible without any intense feature engineering.

1 INTRODUCTION

For many businesses, a significant part of their retail

sales comes from their e-commerce website.

Furthermore, e-commerce has ramped up during the

pandemic around the world, something which is

expected to become a longer-lasting effect.

Subsequently, the effectiveness of e-commerce

solutions, as it is expressed by increased purchase

conversion rates, has become an important challenge

for successful e-businesses.

We can categorize two different approaches for

increasing purchase conversion rates of an e-

commerce web application. The first aims to mine

and analyze the usage of an e-commerce application

(Moe, 2003) to improve its design (Carmona et al,

2012) or, to personalize web page content or the web

site structure that is rendered to different types of

online shoppers (Ding et al, 2015). The second

https://orcid.org/0000-0003-1373-4022

https://orcid.org/0000-0003-4087-125X

https://orcid.org/0000-0002-4416-4790

https://orcid.org/0000-0001-9847-7410

approach aims to create in real time a model of user

purchasing behavior, and to predict the purchase

intent and probability, cart abandonment likelihood

etc. If such predictions would be done accurately,

then conversion rates could be improved by applying

different marketing and web stimuli, such as offers,

coupons, time-limited price discounts and other

purchase incentives. In real settings, the two different

strategies can co-exist implementing two different

pathways to the same objective. The general

assumption underlying both approaches is that the

collective patterns hidden in users’ navigation paths

on e-commerce applications can be analyzed to make

effective predictions about the next actions and

behavior of the current user and her/his purchase

intent.

In this paper we present a real-time e-commerce

analysis component that makes purchase intent

predictions using only the short-term browsing

252

Diamantaras, K., Salampasis, M., Katsalis, A. and Christantonis, K.

Predicting Shopping Intent of e-Commerce Users using LSTM Recurrent Neural Networks.

DOI: 10.5220/0010554102520259

In Proceedings of the 10th International Conference on Data Science, Technology and Applications (DATA 2021), pages 252-259

ISBN: 978-989-758-521-0

 2021 by SCITEPRESS – Science and Technology Publications, Lda. All r ights reserved

pattern of a user. The motivation for using short-term

browsing patterns only, is that in many e-commerce

applications long-term user profiles do not exist,

because users are first-time visitors, or the application

does not require visitors to login, in order to conform

with users’ privacy concerns or other policies such as

the General Data Protection Regulation (GDPR).

In our method, we use Long Short-Term Memory

(LSTM) Recurrent Neural Networks (RNNs) to

process online shoppers’ behavioral data. RNNs

possess several properties that make them suitable for

sequence learning of user sessions. The last few years

a significant line of research has emerged to explore

whether recommendations in e-commerce web sites

can be viewed as a sequence prediction problem using

RNNs. This research was mostly driven by the

findings that RNN produces models with higher

learning capacity and generalization ability than

Hidden Markov Model (HMM).

In fact, several studies have recently used RNN

for session-based recommendation (SBRS) (e.g.

Hidasi and Karatzoglou, 2018; Salampasis et al,

2021). SBRS are recommendation techniques which

rely on the user’s recent actions, the observed

community buying behavior and, other session-

specific data. Generally, in the last five years RNNs

have been widely used in the e-commerce domain to

incorporate temporal features and models for

analyzing sequences of user actions. The main task

was to improve recommendations or, to predict future

behavioral directions. Though, they have been barely

used in the task of predicting purchase intent. From

that perspective, our research work contributes to

better understanding of the RNN-LSTM method for

predicting purchase intent. More specifically we

wished to determine, if and under what e-commerce

scenarios, RNN-LSTM can achieve comparable

results with more traditional ML methods that have

been the state-of-the-art for this task, but they do

require extensive feature engineering processes.

Sakar et al (2019) presented a study which is

much more relevant to our purchase intent task. In

this work they use LSTM-RNN to predict the

probability that the user will leave the site within a

certain time. In our study we use LSTM-RNNs to

predict in real-time the purchase probability, i.e. if the

user will add an item to his or her cart, and if s/he will

finally make a purchase or will abandon the cart.

Also, our approach on creating a model for predicting

purchase intent is different. We utilize all τυπε οφ

actions from previous anonymous user sessions, and

we create a model that in production settings will be

normally used for making predictions after each user

action. If a certain output for the current session is

predicted, for example cart-abandonment, then the e-

commerce application can proactively offer

incentives (e.g. discount coupons, or any other

enticement based on a business or marketing rule)

with a view to increase the probability that a user will

finally make a purchase. This is precisely how we

envisage that an e-commerce system would integrate

our purchase intent component.

To evaluate our method, we use an industry

dataset that has log data from a medium size e-

commerce web application (leather apparel). We also

considered using standard datasets such as the

YooChoose, which was first presented in the RecSys

Challenge 2015. This dataset provides sequences of

click events, product view and buying events, and the

goal is to predict whether the user (a session) is going

to buy something or not, and if s/he is buying, what

would be the items s/he is going to buy. However, this

dataset is not suitable, because our problem definition

and prediction task are different. First, our aim was

not to consider only the product view and buying

events, but the entire spectrum of many different user

action types. This is done in the same sense that it is

studied in the behavior modelling and evaluation of

web-based information seeking systems (e.g.

Salampasis and Diamantaras, 2002). Second, our task

is not to predict if a session will be a buying session

or not as a whole. Instead, our approach is more

dynamic. Specifically, during a user session, and for

each user action, we predict for the remaining session

section, what would be the final outcome (i.e. browse

only, cart abandonment, purchase).

Besides the basic user action data that our e-

commerce Web site stores in its log files, in our

research work, we additionally calculate and utilize

some extra features. These are: the time spent in each

action (in seconds), how the user has landed in the e-

commerce site (origin: referred by an online

advertisement or not), season (autumn/winter or not),

day (Weekend or not) and working hours (yes/no).

We do not include more features because are

potentially difficult to maintain and recalculate in a

realistic production setting, or features relying on user

profiles. Generally, our feature selection process is

driven by the idea to use only features that are

compatible with a strictly session-based operational

scenario. Specifically, we select features that do not

require any user profile data, and additionally are

straightforward to implement in a real-business

environment, that would require frequent (even daily)

re-training and update of the prediction models.

The rest of the paper is organized in the

following manner. Section 2 presents prior work. In

Section 3 we discuss our dataset, how we model the

Predicting Shopping Intent of e-Commerce Users using LSTM Recurrent Neural Networks

253

shopping intent problem, and finally the method that

we used to provide a solution. In Section 4 we

describe the experiments conducted and we report

and discuss the results. Section 5 concludes the paper,

summarizing the findings and presenting ideas for

future development.

2 PRIOR WORK

A large range of statistical, machine learning and

neural network methods have been applied to predict

the purchase probability in e-commerce applications.

For example, simple Bayes and multilayer perceptron

classifiers have been used to predict whether an e-

commerce visitor is likely to make a purchase or not

(Budnikas, 2015).

Suchacka et al (2015) collected data from an

online bookstore and have applied SVM using many

different variables (23) to classify user sessions as

either browsing or buying sessions. The best SVM

classifier proved to be very effective, with an overall

predictive accuracy of 99% and the probability of

predicting a buying session of almost 95%.

Suchacka and Chodak (2017) continued this work

using association rules and a k-nearest neighbor (k-

NN) classifier to assess the purchase probability in

the online bookstore. They analyzed the Web server

log and extracted core user action data (action type

and time spent) as well as other features such as

session length, total session duration in seconds,

average action time in seconds, origin representing

how the user was referred to the bookstore site,

product categories viewed during the session. They

used simple association rule mining to predict with

good accuracy the purchase probability of the users

and other behavior knowledge for two customer

groups: traditional customers (accuracy 90%) and

innovative customers (accuracy 88%). k-NN method

was equally effective, however it is deemed not

suitable for real-time prediction since it is a lazy-

learning algorithm.

The prediction of an e-commerce user purchase

intent requires the predictions of her/his next actions

and it is naturally a sequence-based problem. For that

reason, Hidden Markov Model (HMM) was used as a

promising solution. Generally, there are a lot of Web

usage mining works that use HMM to predict the next

action of a user or infer about her/his overall

navigation behavior. These works have applied and

tested HMM in several contexts and for different

tasks such as predicting web search success (Ageev et

al, 2011), recommender systems (Aghdam et al,

2015), tourism web sites (Yifan et al, 2013).

A research work more related to our study, is

presented in Ding et al (2015) which uses HMM to

learn real-time shopper intent for optimal page

adaptation. To capture shopper’s behavior in real-

time they model and monitor several cart choices

(exit, no change, remove item, add item, purchase).

Their proposed a model that effectively differentiates

each user according to her or his real-time intent. In a

simulated test it manages to reduce shopping cart

abandonment by 32,4% and improves purchase

conversion by 6,9%, if the e-commerce Web site

initiates optimal page adaptation.

The review in this Section shows that the problem

of purchase intent or session classification in e-

commerce applications has been studied with a large

variety of methods. The RecSys2015 challenge asked

the contenders to predict the set of items purchased in

a session based on the user clicks. The challenge

winners (Romov and Sokolov, 2015) proposed a two-

stage classifier trained by a version of gradient

boosting which is still regarded as the State-of-the-

Art (SotA) for this problem. The first classifier makes

a binary prediction whether there will be at least one

buy in the session or not, while the second classifier

predicts the purchased item set. The method used

many categorical features of the sessions and the

items, such as the time and date of the session or the

click, the number of clicks for each item, etc.

However, this approach makes purchase intent

prediction using all click events of a session. Very

few works have attempted to make dynamic

predictions of user purchase intent during the session

where the next user actions are not yet available. Sheil

et al (2018) used recurrent neural networks (RNN) to

capture both session and dataset-global event

dependencies and relationships for user sessions of

any length. Results on benchmark datasets from the

RecSys15 challenge show that their work performed

very close to the RecSys15 challenge winner. The

main difference is that their method does not require

any domain or dataset specific feature engineering for

all types of sessions.

Sakar et al (2019) present another online shopper

behavior analysis system based on RNN-LSTM. It

has two modules operating in parallel and the

experiment they report used a non-public dataset. The

first module predicts shopping intent, but this module

uses ML classifiers such as Random Forest, support

vector machines and multilayer perceptron. Only the

second module uses RNN-LSTM and predicts the

Web site abandonment likelihood without making a

purchase. Their shopping intent module performs

significantly lower that the SotA. The second module

which predicts Web site abandonment after a short

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

254

action window (3 moves) achieves an accuracy of

almost 75%.

Ling et al (2019) use a full connected long short-

term network (FC-LSTM) for modeling the

interactions between customers, and promotion

channels, as well as the nonlinear sequence

correlations and cumulative effects between

customer's browsing behavior. However, to improve

the performance of the prediction they incorporate

more features of customer profile including purchase

history and demographics.

Considering the realities of the e-commerce

domain and the dynamic nature of buying behavior,

we believe that if comparable performance can be

attained using RNN-LSTM and with less demanding

feature engineering, this advantage should be

considered important. The reason is that in e-

commerce applications any prediction model would

need very frequent updates to address the dynamic

nature of online shopping. For example, in a realistic

business environment, a prediction model should be

frequently updated/re-trained, in order to capture

temporal changes in the shoppers buying behavior

and generally accommodate any temporal aspect of

the e-commerce application.

3 PREDICTING PURCHASE

INTENT USING LSTM-RNN

3.1 Dataset

The raw data for the experiments reported in this

paper are taken from the web server logs of an e-

commerce application for a relatively long period of

time (several months). Those kinds of datasets that

are extracted through real applications, entrain all

possible malfunctions. In our dataset, there were

sessions with hundreds of events which admittedly

belong on the maintenance staff and should be

discarded, while there were events that were spanning

long periods or actions and therefore were not

considered reliable. For that reason, we introduced

some limit criteria on our sessions and events and

discarded everything above those. Indeed, sessions

with more than 250 user actions (events) were

removed and any single action lasting more than 10

minutes was limited to 600 seconds.

After we setup some basic data curation, the log data

were analyzed to identify sessions, session length,

user actions in each session, actions’ related items,

item categories, and time spent in each action.

However, not all these data are used in this study, e.g.

item and item categories are not used because our task

was not to predict the item that will be purchased, but

what is the purchase intent. The user-agent cookie of

each log line was used to identify unique sessions.

Each sequence identified by a unique cookie is a

user’s session and consists of all the actions that the

user has made during a session. The above dataset

was further processed to obtain only the sessions that

contain at least 3 user action sequences. As a result of

this preprocess the final dataset consists of 21.896

sessions that altogether count 258.101 user actions.

The average session size is 11,7 actions and the

Median is 8. The average size of Browsing, Cart

Abandonment and Purchase sessions are 11,05, 18,8

and 19,54 respectively.

Table 1: Types of user actions extracted from the Web server log file of our e-commerce application and their frequencies in

the dataset and in each of the three different session types.

All sessions (21.896)

Purchase sessions

(689)

Browse sessions

(19.902)

Cart Abandonment

sessions (1305)

CATEGORY 141561 54,8% 2760 20,5% 128845 58,5% 9956 40,6%

VIEW PRODUCT 83299 32,3% 3149 23,4% 73639 33,5% 6511 26,5%

HOME 6139 2,4% 316 2,3% 5463 2,5% 360 1,5%

ASK_QUESTION 4027 1,6% 150 1,1% 3398 1,5% 479 2,0%

ORDER 730 0,3% 730 5,4% 0 0,0% 0 0,0%

CONTACT 4264 1,7% 42 0,3% 4089 1,9% 133 0,5%

ADD CART 3380 1,3% 1113 8,3% 0 0,0% 2267 9,2%

VIEW CART 10722 4,2% 5011 37,2% 1228 0,6% 4483 18,3%

SEARCH 69 0,0% 0 0,0% 69 0,0% 0 0,0%

CONCERNED 1172 0,5% 115 0,9% 874 0,4% 183 0,7%

ACCOUNT 2068 0,8% 75 0,6% 1838 0,8% 155 0,6%

RECOMMEND 670 0,3% 5 0,0% 656 0,3% 9 0,0%

258.101 100% 13.466 100% 220.099 100% 24.536 100%

Predicting Shopping Intent of e-Commerce Users using LSTM Recurrent Neural Networks

255

The complete set of actions were retrieved for the

website’s user’s behavior analysis (click-stream

analysis) and all user actions (web page visits) were

classified in twelve basic action types. We believe

that these twelve action types represent virtually all

actions that a user could perform in any e-commerce

application, hence they can be used as a “standard”

set of action types for future studies in the e-

commerce domain that consider a broad spectrum of

potential user actions, i.e. not only product views and

purchase actions that are typically considered in

Recommender Systems research. Table 1 shows all

user action types modelled in our system and what is

their appearance frequency in the dataset categorized

in each of the three different session types. Most of

the action types shown in Table 1 are self-explained

and their semantics are apparent. Action type

CONCEREND means that a user has visited a web

page related to privacy policy, payment security or

product shipping & returns. ASK_QUESTION is

triggered when a user makes an enquire during his

session. Finally, action

RECOMMEND

represents the

action of a user recommending a product to another

person (by sending an email notification).

The 689 sessions that ended in purchases

represent a conversion rate of 3,14%, quite an average

rate for apparel e-commerce applications. In the

90.9% sessions (19.902), users did not have any items

in their shopping cart when they exited, which implies

sessions which were pure browsing. The rest of the

sessions (1305) had items in their shopping cart when

they finished, but never turned into purchases

representing the “cart-abandonment” sessions.

3.2 Method

Our method models, for each user session, all user

actions as a sequence representing the user navigation

during the entire session. Each user action belongs to

one action type from those presented in Table 1.

Additionally, each action has a duration which is

calculated in seconds and is the time that user stays in

this web page before a new user action occurs. For

each user action k during a session, the input sample

for the model training are the N events/actions that

precede event k, with N being a parameter of the

training method (action history window).

It is important to clarify that despite an entire

session can be considered as browse-only, cart-

abandonment and purchase, the training of the

prediction model is not done by taking these session

labels to statically determine the output of all session

steps/actions during a session. Instead, for every

sequenced user action in each session, the output is

calculated based on the remaining actions until the

session ends. So, it is possible, during a session, each

session segment to be labelled with a different output.

In other words, a single event (e.g. make order) which

occurs anywhere in the session, will not uniformly

determine the label of all user actions of the entire

session. This is an important modeling detail which

differentiates our method to other research works on

purchase intent reported in the literature.

To prepare the training data in a suitable and easy

to process format, we use the time (in seconds) that

the user spent in this action, to denote which user

action has occurred in each session step as shown in

Table 2. All the other user actions are denoted with 0.

Additionally, we use four extra bits to represent the

extra features that our model uses (origin, season,

day, working hours). These extra features were

selected due to characteristics of the specific e-

commerce application and because previous studies

have shown that buying rate changes over time

Figure 1: Sliding window running from the start to the end of each session to obtain all user’s session instances for training-

prediction purposes.

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

256

(Romov and Sokolov, 2015). The feature Season

indicates a high-season (autumn/winter for leather

apparel) or not. We know that a shopper makes a

purchase with a higher probability on the weekend

than on the working day. We also know that visiting

the e-commerce site during midday leads to purchases

several times more often that in the night hours.

In this way, each session is modeled as a

sequence of events and each event is represented with

a vector like the ones presented in Table 2. Each item

in this vector, in the first twelve positions, represents

one action type from those illustrated in Table 1,

while the rest four positions correspond to the four

extra features. The action that occurs in each session

step is signified with its duration in seconds (>0) and

all other events are marked with 0 value. For example,

in the data that are presented in Table 2, the user

action (Event 1) was a View-Product action that

lasted 265 seconds, the action 3 was again a View-

Product that lasted 270 seconds, and event k was an

ASK_QUESTION user action that lasted 45 seconds.

Using this modeling of the session data, the

training is straightforward. For each session sequence

a sliding window starting from the first session action,

signifies an instance in the e-commerce user

navigation that is used as a sample for training (Figure

1). The length of the window is fixed (N) and it is a

parameter of the method. In cases when the front of

the sliding window lies within the first N-1 events, i.e.

the available user navigation history steps are still

fewer than the size of window, the extra slots are

padded with fully zeroed events.

Table 2: Modeling the user actions sequence and the extra

features (origin, season, day, working hours). Each training

line corresponds to a single session event (user actions) and

has 12 numbers (each number indicating one user action

illustrated in Table 1), plus 4 numbers representing the extra

features.

Event 1 0 ;265 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;1 ;1 ;0 ;1

…

Event 3 0 ;270 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;1 ;1 ;0 ;1

…

Event

0 ;0 ;0 ;45 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;0 ;1 ;1 ;0 ;1

For each input instance that our method prepares

for training, the corresponding target is computed

based on the remaining part of the session (i.e. the

remaining events). Specifically, for every event E

the

method calculates what the target would be,

considering all the remaining session events, starting

from E

i+1

until the last event of the session. The

structure of the target is a bit-alike 2-digit formation.

Each digit represents one of the two actions of interest

(i.e. add cart and make order). That means that there

are 4 possible output statuses. Two of them, (1,0) and

(0,1) signify the existence of at least one add-cart or

one purchase event, and a third one the existence of

both events (1,1). In case of the complete absence of

these two events in the rest of the remaining session

segment, the target becomes (0,0) which means we

have a browse-only remaining session. Figure 2

shows an overview of the prediction model

architecture.

Figure 2: An illustration of the model architecture. In the

input layer, n denotes the number of features/actions and m

denotes the window (sequence) length. Also, x

denotes the

action length in seconds or 1 or 0 for the binary extra

features. After passing the LSTM/GRU and Sigmoid layer

the output is a vector with two binary elements (0 or 1) each

element representing the add_cart event and make_order

events respectively.

To summarize, in our study, the purchasing

intention model is technically designed as a two-label

classification problem, each label representing the

presence or the absence of the add_cart and/or

make_order events, in the session section which

follows immediately after the input session window.

From an application/task point of view, the final

classification can be summarized towards a single

purchase intent scenario outcome. The following four

scenarios describe the four different outcomes

representing the four different targets in the training

phase. Obviously, the same targets are used in the

testing/prediction phase:

 The user will add item(s) to the cart but will not

make a purchase. This target represents the cart

abandonment scenario.

 The user will add item(s) to the cart and will

make a purchase (purchase scenario).

 Neither of the events add-item and make-order

will occur (browse only).

 The user will make a purchase (event). Note that

this last scenario is possible when an add-cart

Predicting Shopping Intent of e-Commerce Users using LSTM Recurrent Neural Networks

257

event has already occurred in the user

navigation history. Subsequently, in the

remaining events of the session, an order is

completed without an add-cart event.

4 RESULTS

We conducted several tests to find the best

hyperparameter tuning. We used 10-fold stratified

cross-validation and we concluded that using 0.2

Dropout rate, Adam Optimizer and 500 LSTM units

produces the best performance. The results using this

specific parameter tuning and window size N=10 are

shown in Table 3. Binary accuracy is the evaluation

metric we use. We notice that LSTM combined with

a GRU layer give the best performance compared to

only one LSTM layer. This finding is in line with

other research for sequence modeling (Chung et al,

2014) variants such as LSTM and GRU regularly

outperform standard recurrent units. Both models end

up to a Dense Layer with sigmoid function to give the

final probabilities for multilabel prediction.

Table 3: Accuracy Results (Window size=10).

Model Units

Extra features

use

Binary

Accurac

LSTM 500

O 97%

LSTM 500 YES 97.3%

LSTM +

GRU

500 X 300

O 97.2%

LSTM +

GRU

500 X 300 YES 97.6%

The results obtained from our method are better

with other research works using RNN-LSTM. Also

they are very much comparable to the accuracy results

that other SotA methods have achieved in similar tasks.

However, these methods were tested in other datasets,

therefore a direct comparison cannot be done.

We conducted more experiments to test our model

using various window sizes. We also wanted to

experiment particularly with the cart abandonment

sessions. These sessions are of particular interest for

e-commerce applications because the user adds

item(s) in the cart, but s/he does not make a purchase.

Apparently, any e-commerce application would

benefit, if these sessions can be effectively predicted,

as soon as possible during a session. Table 4 presents

the Binary Accuracy results using various window

sizes when: a) all sessions are included, and b) when

only the cart-abandonment sessions are considered.

We can see in Table 4 that the classification of

cart abandonment sessions is significantly less

effective, in comparison our method when all

sessions are considered. Similar findings have been

seen in other studies of the same task due to the nature

of the dataset which contains many Browse only

sessions and very few cart abandonment and purchase

session.

Table 4: Accuracy results as a function of Window size.

Window

size

Binary Accuracy

(all sessions)

Binary Accuracy

(Cart Abandonment

sessions

)

4 94.64 56.40

8 96.35 71.00

11 97.45 78.60

14 97.82 81.17

20 98.15 83.16

25 98.14 82.65

30 98.12 83.30

5 CONCLUSIONS

We presented an RNN-LSTM classification method

and model for predicting users’ purchase intent in e-

commerce applications. Our work on this model was

mainly driven by the need to produce a system for

effective predictions that is easy to maintain, re-train

and update. Our model produces very good results,

achieving performance very close to the SotA

methods reported in the literature, although a direct

comparison cannot be made because our task is not

modeled in the same way. However, it is important to

point out that we manage this performance without

using any application explicit features, or excessive

feature engineering.

The model is rather straightforward to implement,

and there is nothing foreseeable to prevent it from

easily generalizing to different datasets and e-

commerce applications with similar performance.

Also, it can be trained with modest hardware resource

requirements and can also provide predictions in real-

time.

Our work was mainly inspired by the idea that e-

commerce web applications should have components

for continuously monitor users during their

navigation. We believe that the primary features such

components should incorporate into the e-commerce

application are: customer short history, collective

experience from community purchase behavior and, a

set of proactive stimuli actions offering buying

incentives to the user. All these features should be

integrated into an end-to-end framework that can be

deployed cost-effectively in small and medium size e-

commerce applications.

DATA 2021 - 10th International Conference on Data Science, Technology and Applications

258

We believe that we have demonstrated the

feasibility of producing such a framework. However,

there are aspects of the framework that we wish to

explore and develop further.

In conclusion then we feel that in this paper we

have already demonstrated a dynamic method based

on RNN-LSTM for effectively predicting purchase

behavior in e-commerce. This method could become

the starting point for developing more complex

frameworks for e-commerce applications that will

aim at higher conversion rates and better profitability.

ACKNOWLEDGEMENTS

This research has been co-financed by the European

Regional Development Fund of the European Union

and Greek national funds through the Operational

Program Co, uder the call RESEARCH–CREATE–

INNOVATE (project code: T1EDK-01776).

REFERENCES

Mikhail Ageev, Qi Guo, Dmitry Lagun, and Eugene

Agichtein. 2011. Find it if you can: a game for modeling

different types of web search success using interaction

data. In Proceedings of the 34th international ACM

SIGIR conference on Research and development in

Information Retrieval (SIGIR '11). Association for

Computing Machinery, New York, NY, USA, 345–354.

Mehdi Hosseinzadeh Aghdam, Negar Hariri, Bamshad

Mobasher, and Robin Burke. 2015. Adapting

Recommendations to Contextual Changes Using

Hierarchical Hidden Markov Models. In Proceedings of

the 9th ACM Conference on Recommender Systems

(RecSys '15). Association for Computing Machinery,

New York, NY, USA, 241–244.

Hidasi Balázs and Karatzoglou Alexandros. Recurrent

Neural Networks with Top-k Gains for Session-based

Recommendations. In Proceedings of the 27th ACM

International Conference on Information and

Knowledge Management (CIKM ’18). Association for

Computing Machinery, New York, NY, USA, 2018,

843–852 (2018).

Carmona CJ, Ramırez-Gallego S, Torres F, Bernal E, del

Jesus MJ, Garcıa S (2012) Web usage mining to

improve the design of an e-commerce website:

OrOliveSur. com. Expert System with Applications

39(12):11243–11249

Ding AW, Li S, Chatterjee P (2015) Learning user real-time

intent for optimal dynamic web page transformation.

Inf Syst Res 26(2):339–359.

C. Ling, T. Zhang and Y. Chen, "Customer Purchase Intent

Prediction Under Online Multi-Channel Promotion: A

Feature-Combined Deep Learning Framework," in

IEEE Access, vol. 7, pp. 112963-112976, 2019

Wendy W. Moe, Buying, Searching, or Browsing:

Differentiating Between Online Shoppers Using In-

Store Navigational Clickstream, Journal of Consumer

Psychology, Volume 13, Issues 1–2, Pages 29-39,

2003.

C Okan Sakar; Polat, S Olcay; Katircioglu, Mete; Kastro,

Yomi.Neural Computing & Applications; Heidelberg

Vol. 31, Iss. 10, (Oct 2019): 6893-6908.

GraźYna Suchacka and Grzegorz Chodak. 2017. Using

association rules to assess purchase probability in

online stores. Inf. Syst. E-bus. Manag. 15, 3 (August

2017), 751–780.

Suchacka, G., M. Skolimowska-Kulig and A. Potempa.

“Classification Of E-Customer Sessions Based On

Support Vector Machine.” ECMS (2015).

Germanas Budnikas. "Computerised Recommendations On

E-Transaction Finalisation By Means Of Machine

Learning," Statistics in Transition New Series, Polish

Statistical Association, vol. 16(2), pages 309-322, June,

2015.

Peter Romov and Evgeny Sokolov. 2015. RecSys

Challenge 2015: ensemble learning with categorical

features. In Proceedings of the 2015 International ACM

Recommender Systems Challenge (RecSys '15

Challenge). Association for Computing Machinery,

New York, NY, USA, Article 1, 1–4.

Michail Salampasis, Theodosios Siomos, Alkiviadis

Katsalis, Konstantinos Diamantaras, Konstantinos

Christantonis, Marina Delianidi and Pantelis

Kaplanoglou. Comparison of RNN and Embeddings

Methods for Next-item and Last-basket Session-based

Recommendations. ICMLC 2021, Shenzhen, China,

February 2021.

Michail Salampasis & Konstantinos Diamantaras. Rich

interactions in digital libraries: Short review and an

experimental user-centered evaluation of an Open

Hypermedia System and a World Wide Web

information-seeking environment. Journal of Digital

Information, BCS & Oxford University Press, ISSN

1368-7506, May 2002.

Humphrey Sheil, Omer Rana, Ronan Reilly. Predicting

Purchasing Intent: Automatic Feature Learning using

Recurrent Neural Networks. eCOM@SIGIR 2018.

Junyoung Chung, Çaglar Gülçehre, KyungHyun Cho, and

Yoshua Bengio. 2014. Empirical Evaluation of Gated

Recurrent Neural Networks on Sequence Modeling.

CoRR abs/1412.3555 (2014). arXiv:1412.3555

http://arxiv.org/abs/1412.3555

Amy Wenxuan Ding, Shibo Li, and Patrali Chatterjee.

2015. Learning User Real-Time Intent for Optimal

Dynamic Web Page Transformation. Info. Sys.

Research 26, 2 (June 2015), 339–359.

Shi, Yifan & Wen, Yimin & Fan, Zhigang & Miao, Yuqing.

(2013). Predicting the Next Scenic Spot a User Will

Browse on a Tourism Website Based on Markov

Prediction Model. Proceedings - International

Conference on Tools with Artificial Intelligence,

ICTAI. 195-200. 10.1109/ICTAI.2013.38.

Predicting Shopping Intent of e-Commerce Users using LSTM Recurrent Neural Networks

259