Neural Comput & Applic (2012) 21:1017–1029

DOI 10.1007/s00521-011-0674-7

ORIGINAL ARTICLE

A framework for improving e-commerce websites usability

using a hybrid genetic algorithm and neural network system
Babak Sohrabi • Payam Mahmoudian •

Iman Raeesi

Received: 12 July 2010 / Accepted: 14 June 2011 / Published online: 8 July 2011
 Springer-Verlag London Limited 2011

Abstract An ever-increasing use of Internet has greatly company, product, or service and further support transac-
been affecting the importance of e-commerce websites. As tion functions for online trading to serve users. The over-
such, analyzing those websites and discovering customers’ loading information would make users feel lost and
purchasing power, the present study tends to propose a frustrated when they surf on the websites. In general, users
framework and an appropriate structure to make the web- prefer and are more comfortable with websites that present
sites more flexible and highly functional. Such an efficient the right content in ways that their linings with preference
and effective framework facilitates customers’ purchasing [6]. Hence, not only finding an appropriate product
capacities. In this framework, different techniques and becomes time-consuming but it also reduces the benefit of
methods including the genetic algorithm, neural networks, the online services [69].
and collaborative filtering are utilized. The pre- and post- To some extent, the problem is being minimized by web
execution data show that the proposed framework changes personalization that aims to ‘‘provide users with the
the website structure based on the two criteria, i.e., sales information they want, without expecting them to ask for
conversion rate and average pre-purchase page views and it’s explicitly.’’ [43] To identify customers’ needs and
finally leading to improved website usability. measure their satisfactions, various methods have been
presented [4, 39]. In customer relationship management,
Keywords Genetic algorithm  Neural networks  the knowledge extraction techniques are applied once there
Website usability  Personalization  Recommender is a large volume of customer data, features, and behavior
systems  e-commerce  Website structure available. Under such circumstances, utilizing algorithms
and efficient techniques becomes essential. The Internet
and e-commerce sites within it contain such data [23]. In
1 Introduction the past few years, the growing use of the Internet has
greatly affected the importance of e-shopping and e-busi-
The continuous growth in the size and the use of the World nesses. These channels make the customers aware about
Wide Web imposes new methods of design and develop- the variety of products and services; hence, they go for the
ment of online information service for most potential best. The more online information available to customers,
market. In order to gain the strategic advantage of further the more market awareness they will have while the more
competition in electronic commerce (EC) on the web, complicated the process of product selection will become
many companies have established their websites as a due to the overwhelming flow of available information.
business frontier [15]. However, most of these websites are Thus, the importance of creating flexible and high-func-
loaded with a large amount of information about the tionality websites that facilitate the delivery of information
and purchase of products and services has increased dra-
matically [16]. Since the managers of e-commerce websites
B. Sohrabi (&)  P. Mahmoudian  I. Raeesi
are looking for ways to eliminate confusion and assist
Department of Information Technology Management,
Faculty of Management, University of Tehran, Tehran, Iran customers in their easy access to the products, there is an
e-mail: bsohrabi@ut.ac.ir evident need for a framework that could improve the

123
1018 Neural Comput & Applic (2012) 21:1017–1029

Fig. 1 The taxonomy of web

mining

structure of their online store, based on customer behavior. information or knowledge from the web hyperlink struc-
The current research aims to propose a framework to dis- ture, page content, and usage data [12]. Data mining by
cover customers’ purchasing behavior and present an automatic or semiautomatic exploration and analysis on a
appropriate structure on the basis of the perceived results. large amount of data items set in a database can discover
Having utilized different techniques and methods including potentially significant patterns inherent in the database
genetic algorithm, neural networks, and recommender [17].
systems, improved website usability was reached. The next Web mining, in general, is the application of data
section deals with the literatures used in this research. mining techniques to discover patterns from the web [37,
Third section proposes a framework. Finally, findings and 45]. For example, using association analysis to analyze
conclusions are discussed in the Sects. 4 and 5. users’ usage data, which records the user’s behavior when
the user browses or makes transactions, on the website and
the results can make the content of the website to fit cor-
2 Literature review rectly with the users’ needs. Different with data mining,
there are no existing data available for web mining. Web
The fast developments on the computer and networking miners can use name or terminology to search and to col-
technologies have increased the popularity of the web lect data [68].
which has caused the inclusion of more and more infor- Different Scholars have proposed taxonomies of web
mation to the web. However, the explosive growth of the mining. Taxonomy of web mining is widely presented in
information on the web has brought some search problems: the literature as shown in Fig. 1 [7, 20, 65]. Web mining is
(1) general purpose search engines often return too many classified roughly into three domains such as web content
irrelevant results when users are searching for specific mining, web structure mining, and web usage mining. Web
information on a given topic and (2) the number of pages to content mining is the process of extracting knowledge from
be indexed by web search systems has been increasing day the content of a number of web documents. Web structure
by day, which makes difficult to keep both automated and mining is the process of inferring knowledge from the
human-maintained indices up to date [46]. In this section, organization and links on the web, while web usage mining
there is an attempt to make use of the strengths of web is automatic discovery of user access patterns from web
mining, collaborative filtering, genetic algorithms, and servers [36].
neural networks in a recommender system so as to provide Web content mining is related with interpreting web
a robust framework for improving the e-commerce web- contents. Most of the web content data are unstructured
sites usability. text data. Web content mining is often viewed as an
intelligent software program that traverses the web to
2.1 Web mining collect information from the websites Lycos, Alta vista,
Web Crawler, etc. which use web content mining tech-
Data mining refers to extracting knowledge from a large nique. As different from the web content mining, web
amount of data [32]. Data mining, also called knowledge structure mining discovers hyperlink relations between
discovery in databases, is commonly defined as a process websites. It is related to links between web pages directly
of discovering useful patterns or knowledge from row data while web content mining is related to web documents.
sources. In other words, data mining is a multidisciplinary Web usage mining is discovered as a useful knowledge by
field that involves information retrieval, machine learning, using data mining techniques from the secondary web data.
statistics, databases, artificial intelligence, and visualiza- Secondary web data include data from web server access
tion. Similarly, the web mining aims to discover useful logs, proxy server logs, browser logs, user profiles,

123
Neural Comput & Applic (2012) 21:1017–1029 1019

registration files, user sessions or transactions, user queries, entrainment etc. are increasing far more rapidly than our
bookmark folders, mouse clicks and scrolls, and any other ability to process it. All of them such as Amazon, Yahoo,
data generated by the interaction between users and the and CDNow suggest a lot of chooses to their potential
web. Logs of web access available on most servers are customers everyday which makes difficult to find the true
good examples of the data sets used in web usage mining. products that best meet user’s needs and preferences. For
Web usage mining has become very critical for effective overcoming this problem, recommender systems appeared
website management, creating adaptive websites, business and became an appealing subject for researchers. It is a
and support services, personalization, network traffic flow kind of personal software assistant learning the evolving
analysis, and so on [65]. interests of their users by applying the information-pro-
By utilizing web mining on raw data stored on the ser- cessing algorithms to the mass of this information [3].
ver, one can access web usage patterns, website subject Recommender system intends to provide people with
priorities, and the most visited sections and pages. This recommendations on products they appreciate based on
information helps in finding ways to attract more visitors. their past preferences, purchase history, and demographic
Also, by applying web usage mining techniques, one can information [51]. Many of the most successful systems
acquire information about users’ navigational patterns. make use of collaborative filtering [29, 59]. In fact, this
Particularly, web usage mining refers to an automatic system utilizes knowledge discovery techniques in order to
discovery and analysis of patterns in click stream and interactively suggest customized products at e-commerce
associated data collected or generated as a result of users’ websites [50]. In other words, the system is rapidly
interactions with web resources [20, 41, 62]. Here, the goal becoming a crucial tool in e-commerce on the web [56].
is to capture, model, and analyze the behavioral patterns Data/web mining is used for improving the quality of
and profiles of users, interacting with a website. While a CF-based recommender systems frequently. For example,
customer is visiting a website he leaves a trail (which Cho and Kim [18] proposed a recommendation method-
characterizes his requirements) in the form of a server log. ology based on web usage mining, and product taxonomy
To enable predictive modeling of the current customer’s to enhance the recommendation quality and the system
navigation behavior, the data mining provides the facilities performance of current CF-based recommender systems.
for automated discovery of the knowledge within these Recommender systems apply data analysis techniques to
logs [5]. problem of helping users find the items they would like to
Following the standard data mining process [22], the purchase at e-commerce or entertainment sites by produc-
overall web usage mining process can be divided into three ing a predict likeliness score or a list of top-N recom-
interdependent stages: pre-processing that normally con- mended items for a given user. Collaborative filtering is the
sists of removing useless and excessive data from the raw most important personalized recommendation method
dataset, recycling data related to website structure and web widely used in recommender systems.
pages, and also formatting raw study data [19]. Pattern Recommender systems for large e-commerce sites have
discovery: In this stage, resultant data from the previous to deal with millions of customers and products. Because
stage are analyzed and their behavior pattern is discovered. these systems usually handle very high dimensional pro-
The discovery of rules, conditions and statistical relations files to form the neighborhood, the nearest neighbor
between data are considered as the outputs of this stage. algorithm is often very time-consuming and scales poorly
Pattern analysis: The identified pattern and rules from the in practice. To address the scalability problems in CF-
previous stage are analyzed, and their application for based recommender systems, a variety of approaches have
problem solving is examined. These patterns help analyze been developed. These approaches can be classified into
customer activities within the website, identify their needs, two main categories: dimensionality reduction techniques
and make appropriate changes by improving interactivity and model-based approaches [3, 54]. Latent Semantic
between customer and website. The results accrued from Index (LSI) is a widely used dimensionality reduction
the mining of web logs can also be used (1) to personalize technique. It uses singular value decomposition (SVD) to
the presentation of web contents; (2) to improve user factor the original rating space into three matrices and
navigation; (3) to improve web design or e-commerce sites; performs the dimensionality reduction by reducing the
and (4) to improve the customers’ satisfactions [1]. singular matrix. In model-based approaches, a model is
first built based on the rating matrix and then the model
2.2 Recommender systems is used in making recommendations. Usually, the model is
expensive to build, but rapid to execute. Several data
The amount of information in the world is growing so mining techniques such as Bayesian network, clustering
quickly with the widespread and easy usage of Internet. and association rule mining have been applied to building
Therefore, acquired data from the sites of e-commerce, the model [18].

123
1020 Neural Comput & Applic (2012) 21:1017–1029

The tremendous growth of products and customers in profile may be used to rate other unseen information
recent years as well as high-quality recommendations has sources or to construct a query of a search engine. In
raised serious challenge for recommender system. For contrast to content-based techniques, collaborative meth-
solving this challenging problem, singular value decom- ods do not need any information about item’s content.
position (SVD)-based recommendation algorithms can The task in collaborative filtering is to predict the utility
quickly produce high-quality recommendations. of items to a particular user (the active user), based on a
database of user votes from a sample or population of
2.2.1 SVD algorithm other users (the user database). Both approaches share the
common aim of assisting in the user’s search for items of
SVD, which is the algorithm used in this study, is a matrix interest. Consequently, collaborative filtering (CF) is used
factorization technique that takes an m 9 n matrix A, with as an important approach to classify users and predict
rank r, and decomposes it as follows: their needs in the web domain. It is mostly used for
SVDðAÞ ¼ U  S  V T : providing the customers with categories of products and
services that have been previously provided to the cus-
U and V are orthogonal matrices with dimensions m 9 m tomers with similar purchases and attitudes toward
and n 9 n, respectively. S, called the singular matrix, is an e-commerce website offerings.
m 9 n diagonal matrix whose diagonal entries are non- In the proposed system, the collaborative filtering
negative real numbers. SVD can provide the best low-rank application uses k-nearest neighbor kNN classifiers to
approximation of the original matrix. Vk is produced by predict users’ ratings or purchase propensity by measuring
removing t 9 k rows from matrix V. Matrix Ak is defined as the correlations between a current (target) and past users’
[66]: profiles in order to find their similar characteristics or
preferences [30, 31]. A profile may consist of a set of item
Ak ¼ Uk  Sk  VkT :
ratings or a set of items visited or purchased products.
Ak represents the closest linear approximation of the ori- Essentially, the kNN approach-based collaborative filtering
ginal matrix A with reduced rank k. Once this transfor- consists of a comparison between activity record for the
mation is completed, users and items can be thought of as targeted users and the historical records of other registered
points in the k-dimensional space [66]. SVD, as part of users, T, in order to find the top k users with the most
Latent Semantic Indexing, was used widely in the area similarities in terms of interest and preferences. Predomi-
of information retrieval in order to solve the problems of nantly, in most collaborative filtering tools, activity records
synonymy and polysemy [21]. SVD requires a complete or users’ profiles consist of a collection of their submitted
matrix to operate. Billsus and Pazzani [11] initially utilized ratings about a sub-category of products. Identifying
SVD in order to formulate collaborative filtering as a neighbors, user’s active products can be identified that he
classification problem. Sarwar et al. [55] proposed a has never seen or purchased [40, 48].
method using average values in the empty cells of the The item-based collaborative filtering [57] is in accor-
rating matrix. Another alternate method by [61] finds a dance with the rating pattern among users. In this method,
model that maximizes the log-likelihood of the actual rat- the kNN algorithm tries to find the similar k with regard to
ings by an EM procedure. Furthermore, techniques like the similarity of users’ ratings for the same product.
SVD-updating or folding-in were proposed to alleviate the Applying the following equation, the amount of similarity
problem of updating, which refers to the process of adding between the two products is calculated by the adjusted
new terms and/or documents to existing matrices [10]. The cosine similarity method. Collaborative filtering methods
incremental folding-in SVD method is used in this research work on a table of U users who can rate I items. The
since it is proven to be an efficient method for highly prediction of a non-rated item i for a user u is computed
scalable recommender systems [21, 55, 66]. as an aggregate of the ratings of the K most similar users
(k-neighborhoods) for the same item i, where Ku denotes
2.3 Collaborative filtering the set of k-neighborhoods of u and rn,i denotes of value of
the user n rating on the item I [3, 13].
There are mainly two distinct techniques by today’s rec- Once the set of K users (neighborhoods) similar to active
ommendation systems: content-based methods and col- u has been calculated, in order to obtain the prediction of
laborative methods [3]. Content-based methods analyze item i on user u, one of the following aggregation
the content of information sources (e.g., the HTML approaches is often used: the average (2), the weighted sum
source of web pages) that have been rated to create a (3), and the adjusted weighted aggregation (deviation-
profile of the user’s interests in terms of regularities in the from-mean) (4). We will use the auxiliary set Gu,i in order
content of the information that was rated highly. This to define (2–5) [3, 13]:

123
Neural Comput & Applic (2012) 21:1017–1029 1021

  
Gu;i ¼ n 2 Ku 9rn;i 6¼  ; ð1Þ Recommender systems can be considered as social net-
working tools that provide dynamic and collaborative com-
1 X munication, interaction, and knowledge. The recommender
pu;i ¼ rn;i , Gu;i 6¼ ;; ð2Þ
#Gu;i n2G systems stage that normally has the greatest influence on the
u;i
X quality of the results obtained is the collaborative filtering
pu;i ¼ lu;i simðu; nÞrn;i , Gu;i 6¼ ;; ð3Þ phase [13, 53]. The current study, therefore, has employed an
n2Gu;i
X item-based collaborative filtering that can suggest similar
pu;i ¼ ru þ lu;i simðu; nÞ ðrn;i  ru Þ , Gu;i 6¼ ;; ð4Þ products to those the customer is viewing in an e-commerce
n2Gu;i website.
where l serves as a normalizing factor, usually computed:
, 2.4 Artificial neural networks
X
lu;i ¼ 1 simðu; nÞ , Gu;i 6¼ ;: ð5Þ
n2Gu;i
Artificial neural networks have been developed as gener-
alizations of mathematical models of biological nervous
The most popular similarity metrics are Pearson correlation systems [35]. Artificial neural network is a functional
(6), cosine (7), constrained Pearson’s correlation (8), and abstraction of the biologic neural structures of the central
Spearman rank correlation (9): nervous system [24]. They are powerful pattern recognizers
P and classifiers. They operate as black box, model-free, and
i ðrx;i  r x Þðry;i  ry Þ
simðx; yÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P P ; ð6Þ adaptive tools to capture and learn significant structures in
i ðrx;i  r x Þ2 i ðry;i  ry Þ2 data. Their computing abilities have been proven in the
P fields of prediction and estimation, pattern recognition, and
i rx;i ry;i
simðx; yÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffi
P 2 qffiffiffiffiffiffiffiffiffiffiffiffi
P 2ffi; ð7Þ optimization [25, 26]. They are suitable particularly for
r
i x;i r
i y;i problems too complex to be modeled and solved by clas-
P sical mathematics and traditional procedures [24].
i ðrx;i  rmed Þðry;i  rmed Þ Abraham [2] introduces various ingredients of natural
simðx; yÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ; ð8Þ
2 P 2
ðr
i x;i  r med Þ ðr
i y;i  r med Þ computation and presents the important concepts of web
usage mining and its various practical applications. The gen-
rmed: median value in the rating scale, eral area of computational intelligence is currently undergoing
P an important transformation by trying to incorporate compu-
i ðrankx;i  rankx Þ ðranky;i  ranky Þ
simðx; yÞ ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
P  2 P  : tational ideas borrowed from the nature all around us like
i rank x;i  rank x i rank y;i  rank y Artificial immune system/Swarm intelligence/Evolutionary
ð9Þ computation/Neural computation/Bacterial foraging/DNA
computation/Computing with words/Simulated annealing/
Although Pearson correlation is the most commonly Cultural algorithms/Membrane computing/Quantum compu-
used metric in the process of memory-based CF (user to tation/Artificial life/Hybrid approaches.
user), this choice is not always backed by the nature and In a simplified mathematical model of the neuron, the
distribution of the data in the RS. Formally, in order to be effects of the synapses are represented by weights that
able to apply this metric with guarantees, the following modulate the effect of the associated input signals, and the
assumptions must be met: Linear relationship between x non-linear characteristic exhibited by neurons is repre-
and y; Continuous random variables; and Both variables sented by a transfer function, which is usually the Sigmoid,
must be normally distributed. Gaussian function etc. To develop computational tools for
These conditions are not normally met in real RS, and solving engineering problems, the neuron impulse is then
Pearson correlation presents some significant cases of computed as the weighted sum of the input order. In neural
erroneous operation that should not be ignored in RS. network literature, there are several network topologies,
Despite the deficiencies of Pearson correlation, this simi- each with its own capabilities. The feed-forward perceptron
larity measure presents the best prediction and recom- model is one such, appropriated during the present study.
mendation results in CF-based RS. Furthermore, it is the In the feed-forward networks, the inputs from the interface
most commonly used and therefore any alternative metric layer are connected to the nodes of subsequent layer which
proposed must improve its results [13]. themselves are connected to either the next layer or the
Recommender systems (RS) cover an important field output layer. Back propagation training algorithm involves
within collaborative services that are developed in the Web calculating the derivatives of the network training error
2.0 environment [13, 33] and enable user-generated opinions with respect to the weights by the application of chain-rule
to be exploited in a sophisticated and powerful way. and gradient descent optimization to adjust the weights to

123
1022 Neural Comput & Applic (2012) 21:1017–1029

minimize the error. Neural networks have principle appli- sort the entire population by descending fitness values.
cations in pattern recognition, clustering, forecasting, Once the selection process has finished, it comes to the
extrapolation, etc. [34, 64]. reproduction stage where this generation performs repeat-
edly until one of the termination criteria is met. Zang et al.
2.5 Genetic algorithms [70] have listed a number of the most successful
applications.
Based on Charles Darwin’s theory of evolution, the genetic GA has been widely adopted to solve complex prob-
algorithm is one of the random searching methods initially lems, especially in the areas of scheduling, global optimi-
invented in 1975 by Holland [28]. These algorithms present zation, and control engineering. The implementation of
solutions to a certain problem (which may be random) in genetic algorithms is one of the most utilized methods
order to produce better results [27]. Finding optimum order compared with alternative intelligent optimization algo-
of web pages is an important issue for web designers [47]. rithms [70].
Asllani and Lari [8] employed genetic algorithms in their The biggest difference with other metaheuristics [like
research to achieve the optimum design of a website based tabu search (TS) or simulated annealing (SA)] is that GA
on multiple optimization indicators such as minimization maintains a population of solutions rather than a unique
of download time, maximization of website visualization, current solution. Solutions are coded as finite-length strings
and product association level. They also tried to introduce a called chromosomes, and a measure of their adaptation (the
genetic algorithm to the web design teams which led to an fitness) is computed by an engine. Starting from an existing
increase in sales, customer returns, and user-friendly population, each iteration generates new chromosomes by
website visualizations. Although the design of a genetic applying operators (like crossover and mutation) to two
algorithm is based on other presented methods [44, 49, 52], chosen parents. The main advantage of GA is its intrinsic
the main problem other methods encounter is that they fail parallelism, which allows the exploration of a larger
to consider customer preferences. solution space [58]. A well-designed GA allows for the
GA generates solutions to search, optimization and efficient and effective exploration and exploitation of the
machine learning problems via applying techniques problem’s search space of feasible solutions in an effort to
inspired by biological evolution. GA adopts some genetic identify the global optima, or near-optimal, solution to
terminologies, including [46, 70]: difficult problems [42].
1. Chromosome is an encoding of a solution to an
optimization problem. The solutions are typically
represented in binary [14]. 3 Proposed framework
2. Selection is a stage in the GA where individual
genomes are chosen for breeding new generation. The websites studied in this research consist of different
3. Crossover and mutation both are genetic operators sections displaying and categorizing information related to
applied to a pair of parents when they reproduce to books. Since the aim is to develop a framework to cus-
alter their genetic composition [63]. tomize the layout and categories based on customer
behavior and ratings, such customization will result in an
Overall, the basic steps of GA are as follows [9]:
increase in sales conversion rates and decrease in pre-
purchase page views.
To predict customer behavior on a website, firstly, a
mathematical model needs to be designed. With regard to
the numerous capabilities of mathematical graphs, this
research has considered the website structure using those
graphs. In this particular model, nodes and crest lines
represent web pages and internal webpage links,
respectively.
To perform a GA, the first step is to initialize the pop-
The stages of the proposed framework are as follows:
ulation that is normally composed of randomly generated
individuals covering the entire range of possible solutions, • Data collection and pre-processing;
and size of the population is determined by the nature of • Discovery and analysis of web usage patterns;
the problem itself. The next step is to evaluate the fitness of • Implementation of the SVD algorithm on user ratings;
each member of the population. A fitness function is • Modeling the website structure;
employed in this stage to provide fitness values for each • Learning user navigational patterns through neural
individual, and the results are then normalized in order to networks;

123
Neural Comput & Applic (2012) 21:1017–1029 1023

• Utilizing the genetic algorithm in order to improve 3.2.1 User identification

website usability.
The analysis of web usage does not require knowledge
about a user’s identity; however, he needs to be distin-
3.1 Data collection and pre-processing
guished among different users. Since a user may visit a
site more than once, the server logs record multiple
Input data of the proposed framework consist of two parts:
sessions for each user. Bing [12] has used a phrase ‘‘user
server logs and web application (e-shop website) logs.
activity record’’ to refer to the sequence of logged
With a slight review of this server log as well as the
activities of the same user. Usage processing is arguably
website content structure, we could notice that a user views
the most difficult task in web usage mining process due
‘‘Arts and Photography’’ category when a book entitled
to the incompleteness of the available data unless a
Playing for Pizza: a Novel catches his attention; never-
client side tracking mechanism is used. Only the IP
theless, among all the suggested books he selects the book
address, agent, and server side click stream are available
with ISBN ‘‘0446577391’’ (Fig. 2).
to identify users and server sessions. Therefore, the
present study has used the IP ? Agent user identification
3.2 Discovery and analysis of web usage patterns
method [19]. Assuming the users are identified (through
cookies, logins or IP/Agent/Path Analysis), the click
After collecting and refining the raw server data, the web
stream for each user must be divided into sessions (See
usage patterns that represent the users’ behavior on the
Fig. 3). In the next session, the process of sessionization
website are extracted in three different stages, i.e., user
is explained.
Identification, sessionization and path completion, which
needs further clarifications.

Fig. 2 An example of the

server log [38]

Fig. 3 Information gathered

from the e-shop application

123
1024 Neural Comput & Applic (2012) 21:1017–1029

3.2.2 Sessionization Further, at this stage only, the SVD algorithm is

implemented on the standardized data. The amount of
Sessionization is a process of segmenting the activity of users’ interest in books is calculated on the basis of users’
each user into sessions, each representing a single visit to submitted ratings and their similarities. Here, the output of
the site. A sessionization heuristic aims to reconstruct, this stage determines how interested each user in each book
from the click stream data, the actual sequence of actions is. The data are displayed as a matrix where columns and
performed by one user during his single visit to the site. rows indicate books and the amount of users’ interest in a
Various heuristics for sessionization have been identified specific book, respectively. A part of the implemented code
and studied [60], but the current research uses a time-ori- for SVD algorithm is present here:
ented heuristic method for data sessionization.

3.2.3 Path completion

This is yet another potentially important task after session-

ization. Path completion is a stage when an actual user
behavior pattern is extracted. For example, paths related to
the Fig. 4 ‘‘Session 1’’ are extracted as A*B and A*C*E.

3.3 Implementation of the SVD algorithm on user ratings

For making genetic mutation intelligent, this research has

applied the SVD recommender system. In order to use
information extracted from the users’ ratings database,
firstly, these are converted into a standard form. The input
of this stage is the matrix of users’ ratings. At this stage, the
book numbers that are in ISBN form and acquired from the
database are standardized (from zero to the overall number
of books minus one). Such transformation should also be
performed on users’ numbers. Meanwhile, the required
input for the SVD program is also produced at this stage.

Fig. 4 An example of user

identification using IP ? agent

123
Neural Comput & Applic (2012) 21:1017–1029 1025

A part of the similarity equation is also presented here: and is already on page k. Consequently, the network input is
k coded and weighted clicks made by the user. Here, the input
layer has two nodes for each web page: one displays the pages
a user has visited, and the other shows the existing links on the
current page. In the output layer, there is only single node for
each page. A part of the implemented code is presented here:

3.4 Modeling the website structure

In this stage, previous data are converted into higher level

of information including:
• Adjacency matrix of the graph of web pages. In this case,
when a member is equal to one (1), it means the pages
relevant to that row and column are linked or vice versa.
• Information related to each user’s interest in each book
category. The amount of interest in each category is calculated
using his average interest in the books of that category.

3.5 Learning user navigational patterns through neural

networks

In the current study, the applied neural network is a multi-

layered perceptron network along with back propagation
learning method. A back propagation network (BPN) is a
neural network that uses a supervised learning method and
feed-forward architecture. A BPN is one of the most fre-
quently utilized neural network techniques for classification
and prediction [67] and is considered an advanced multiple 3.6 Utilizing the genetic algorithm in order to improve
regression analysis that can accommodate complex and non- website usability
linear data relationships [17]. The back propagation network
architecture is illustrated in Fig. 5: This network ably gives the In this stage, the genetic algorithm is implemented on web
probability of choosing any of the links available on the page structure to improve the usability of website. The following
k, with the consideration that the user has viewed k - 1 pages algorithm is a pseudo-code of the implemented genetic
algorithm:

Fig. 5 The back propagation network architecture

123
1026 Neural Comput & Applic (2012) 21:1017–1029

Crossover, mutation, and selection are three steps 3. Selection: This is a process by which selection and
repeatedly performed and are illustrated below. transfer from one to the next generation are performed.
Population size in this step is (1 ? c) * p (i.e., initial
1. Crossover: This function performs the crossover oper-
population size plus generated offspring in crossover
ation in the genetic algorithm. For this, first, c% of the
section). Here, using roulette wheel selection, p members
population is selected randomly (Here, c is set to 60 by
from population are selected to be in the next generation.
default) and then, the selected items are placed in a sub-
Interestingly, in roulette wheel selection, each member
population category. Thereafter, each step randomly
has the chance to be selected equal to its fitness value
selects two members and created two offspring through
compared to total fitness value of total population.
the crossover operation. Here, a universal crossover for
graphs, i.e., each offspring has vertex set like its parents,
and for each edge (u, v), the two offspring have the same
chance of inheriting it from mom or dad. 4 Results and discussion
2. Mutation: The mutation operation is performed in the
genetic algorithm on m% of sub-population resulted After the design and implementation, the proposed framework
from the crossover operator. Considering that genetic has been assessed in the bargainsate.com website in the time
algorithms are time-consuming, attempts are usually period of 2007–2008. The log file was processed after the
made for the intelligent mutation. Having the amount completion of the system coding and testing. With reference to
of similarity between each pair of books, the genetic the contract between the researcher and Bargain State man-
mutation operator can be made more intelligent by agement, the managers considered two criteria to assess the
selecting the books with more similarity to each user’s improvement, i.e., the increase in sales conversion rate and
interests. For the mutation, firstly, one member of the decrease in pre-purchase average page views. With the
previous generation is randomly selected and then the coordination of the company’s technical team, a user behav-
links of that member are altered through the following ior-based shopper agent was created to simulate customer
four ways. It must be noted that users’ interests, at this behavior on the website. The agent selects a person from the
state, are calculated by the SVD method that was registered users in Bargain State’s web-based e-shop appli-
explained in Sects. 2.2 and 2.4. cation, and then considering the numerical value of the users’
ratings to books (the amount of user interest); he begins to
• Eliminating links between books and sub-catego-
browse the web pages. Each page can be one of the three pages
ries: To eliminate a link for e1 times, first, a user is
related to the books, their categories or Add to Cart links. The
randomly selected, and then a category he has the
chance of selecting each category among different books is
least interest in is selected. Among the books of the
equal to the average of user ratings within that category.
selected category, the one with the least member-
Meanwhile, the chance of purchasing a book by a user is equal
ship degree is eliminated.
to the multiple of the SHOP-RATE constant and the numerical
• Eliminating links between similar books: To elim-
value of the user’s rating to that specific book. The results of
inate a link for e2 times, a user is randomly chosen,
simulating user behavior through the above process show that
and then the book he has the least interest in is
the sales conversion rate has increased and the average pre-
selected. Among the books, the least similar one
purchase page views have decreased.
based on similarity relation is selected. The link
Following is output of the shopper agent before exe-
between two books is eliminated, if the amount of
cuting the proposed framework’s procedures:
similarity is more than the minimum threshold.
• Adding links between books and sub-categories: To
add a book to a sub-category, the following actions
are performed for a1 times: First, a category is Conversion rate: 3.87%
randomly selected, and then a book with the Average clicks: 10.64
highest degree of membership to this category is
selected and added to this category.
• Adding links between similar books: To add a book Following is the output of the shopper agent after execut-
to a sub-category, the following actions are ing the proposed framework:
performed for a2 times: First, a book is randomly
selected (more likely from books that do not have
similar books or have a few similar category).
Conversion rate: 4.13%
Thereafter, a book is selected among those that are
not similar to that one. Average clicks: 8.56

123
Neural Comput & Applic (2012) 21:1017–1029 1027

Fig. 6 An example of
sessionization with a time-
oriented heuristic

The assessment shows that user was able to find and pur-
chase his desired books faster. Figure 6 highlights the
numerical differences between the two above criteria
before and after implementing the proposed framework
(Fig. 7).
The acquired results show that sales conversion rate
increased by 6.71% while the average pre-purchase page
views decreased by 19.55%. As such, incorporating the
proposed framework and the tools accompanying it, com-
panies can create a more efficient website structure that
leads to an increase in customer satisfaction as well as
profitability. Compare to past efforts, the present study is
more able to deal with various situations effectively in two
ways:First, to achieve an appropriate solution, it merges
Fig. 7 Before–after comparison chart
computer science and management. On this basis, various
researches in different fields of web mining such as genetic
algorithm, neural networks, and recommender system are customized delivery of goods and services. Instead of
examined. conventional shopping, EC provides alternative ways for
Second, the study presents an integrated framework users to get information on products such as price, avail-
through which the desired objectives could be achieved. ability, suppliers, substitutes, and even manufacturing
Within this basis, after reviewing literatures in different process. The overwhelming amount of information about
fields as well as combining them, a framework was created products and services has led to relatively more confusion
through which the desired objectives like an increase in the on whether to purchase a product or to search more for a
rate of sales conversion and a decrease in pre-purchase better result.
page views could be achieved. Considering the ever- In general, there are two major approaches to provide
increasing number of companies that provide products and personalized information: content-based and collaborative
services through the Internet, the need for effective tools to filtering. In the collaborative filtering approach, it identifies
improve website efficiency and save customer time, other users who have showed similar preference to the
through guiding him to relevant products and services, is given users and provides what they would like.
evident. Such efforts can persuade the customers to vastly Nowadays, most recommendation systems consider user
used online markets. preferences, interests, and browsing behaviors when ana-
lyzing the behaviors for personalized services. In this
study, there has been an attempt to provide the scholars and
5 Conclusion and further research practitioners with a novel framework for improving the
e-commerce websites usability using a hybrid genetic
A decision maker’s preference is usually determined by algorithm and neural network approach for implementing a
multiple decision attributes. Normally, the decision maker learning system which seeks for the best possible solution
aims to maximize his utility function that depends on these for the customers. Based on the findings, it can be implied
criteria. Electronic commerce has been widely used by that the proposed framework could be used to build a
online users to perform different daily activities. EC helps website structure which is able to search through the pat-
merchants and consumers to reduce costs and enables terns of users’ website usage in order to better provide

123
1028 Neural Comput & Applic (2012) 21:1017–1029

them with sophisticated and pleasant alternatives. By this 11. Billsus D, Pazzani MJ (1998) Learning collaborative information
framework, not only users will be satisfied, but e-shops’ filters. 15th International conference on machine learning, Mad-
ison, WI, pp 46–53
earnings will increase in an incremental but satisfactory 12. Bing L (2008) Web data mining: exploring hyperlinks, contents,
trend. The current research has indicated that the benefits and usage data. Data-centric systems and applications series.
of the website have increased by 6.71% while page views Springer, Berlin
before a buy action decreased by 19.55%. The stages of the 13. Bobadilla J, Serradilla F, Bernal J (2010) A new collaborative
filtering metric that improves the behavior of recommender sys-
proposed framework can be summarized as follows: data tems. Knowl Based Syst 23:520–528
collection and pre-processing from the selected website; 14. Cha SH, Tappert Charles C (2009) A genetic algorithm for
discovery and analysis of users’ web usage patterns; constructing compact binary decision trees. J Pattern Recognit
implementation of the SVD algorithm on user ratings; Res 4:1–13
15. Chang CC, Chen PL, Chiu FR, Chen YK (2009) Application of
modeling the website structure; learning user navigational neural networks and Kano’s method to content recommendation
patterns through implementing neural networks; and uti- in web personalization. Expert Syst Appl 36:5310–5316
lizing the genetic algorithm in order to improve the website 16. Chircu AM, Mahajan V (2006) Managing electronic commerce
usability. transaction costs for customer value. Decis Support Syst
42(2):898–914
In future researches, other techniques like hill climbing 17. Chou PH, Li PH, Chen KK, Wua MJ (2010) Integrating web
algorithm can be used instead of genetic algorithm for mining and neural network for personalized e-commerce auto-
producing satisfactory results while comparing the final matic service. Expert Syst Appl 37:2898–2910
outputs with the current study. Also, it is possible to con- 18. Cho YH, Kim JK (2004) Application of web usage mining and
product taxonomy to collaborative recommendations in e-com-
struct multiple site structures for each cluster according to merce. Expert Syst Appl 26:233–246
their preferences. Another way for improving the proposed 19. Cooley R, Mobasher B, Srivastava J (1999) Data preparation for
framework for making it usable for larger websites is to mining World Wide Web browsing patterns. Knowl Inf Syst
cluster the web pages according to their similarity and have 1(1):5–32
20. Cooley R, Mobasher B, Srivastava J (1997) Web mining: infor-
a neural network for each cluster to learn and to predict the mation and pattern discovery on the World Wide Web. In: Pro-
user behaviors. ceedings of the 9th IEEE International conference on tools with
artificial intelligence (ICTAI’97), pp 558–567
21. Deerwester S, Dumais ST, Furnas GW, Landauer TK, Harshman
R (1990) Indexing by latent semantic analysis. J Am Soc Inf Sci
41(6):391–407
References 22. Fayyad UM, Piatetsky-Shapiro G, Smyth P (1996) From data
mining to knowledge discovery: an overview. In: Advances in
1. Abedin B, Sohrabi B (2009) Graph theory application and web knowledge discovery and data mining, American Association for
page ranking for website link structure improvement. Behav Inf Artificial Intelligence (AAAI), MIT Press, pp 1–34
Technol 28(1):63–72 23. Flavian C, Guinaliu M, Gurrea R (2006) The role played by
2. Abraham A (2005) Natural computation for business intelligence perceived usability, satisfaction and consumer trust on website
from web usage mining. Seventh international symposium on loyalty. Inf Manage 43(1):1–14
symbolic and numeric algorithms for scientific computing (SY- 24. Gholizadeh S, Pirmoz A, Attarnejad R (2011) Assessment of load
NASC’05), pp 3–10 carrying capacity of castellated steel beams by neural networks.
3. Acilar AM, Arslan A (2009) A collaborative filtering method J Constr Steel Res 67:770–779
based on artificial immune network. Expert Syst Appl 25. Gholizadeh S, Salajegheh E, Torkzadeh P (2008) Structural
36:8324–8332 optimization with frequency constraints by genetic algorithm
4. Ahn T, Ryu S, Han I (2007) The impact of web quality and using wavelet radial basis function neural networks. J Sound Vib
playfulness on user acceptance of online retailing. Inf Manage 312:316–331
44(3):263–275 26. Gholizadeh S, Salajegheh E (2010) Optimal seismic design of
5. Anderson CR (2002) A machine learning approach to web per- steel structures by an efficient soft computing based algorithm.
sonalization. Ph.D. thesis, University of Washington J Constr Steel Res 66:85–95
6. Aragonees M, Hart-Davidson W (2002) Why, when and how do 27. Goldberg DE (1989) Genetic algorithms in search, optimization,
users customize web portals? In: Proceedings of IPCCs, IEEE, and machine learning. Addison-Wesley, Wokingham
pp 375–388 28. Holland JH (1975) Adaptation in natural and artificial systems.
7. Arotaritei D, Mitra M (2004) Web mining: a survey in the fuzzy University of Michigan Press, Ann Arbor (2nd edn. MIT Press, 1992)
framework. Fuzzy Sets Syst 148:5–19 29. Herlocker JL, Konstan J, Borchers A, Riedl J (1999) An algo-
8. Asllani A, Lari A (2007) Using genetic algorithm for dynamic rithmic framework for performing collaborative filtering. In:
and multiple criteria web-site optimizations. Eur J Oper Res Proceedings of the 22nd International conference on research and
176(3):1767–1777 development in information retrieval (SIGIR’99), Berkeley,
9. Azadeh A, Saberi M, Anvari FM, Azaron A, Mohammadi M California. ACM Press
(2011) An adaptive network based fuzzy inference system— 30. Herlocker JL, Konstan J, Riedl J (2000) Explaining collaborative
genetic algorithm clustering ensemble algorithm for performance filtering recommendations. In: Proceedings of the ACM 2000
assessment and improvement of conventional power plants. conference on computer supported cooperative work
Expert Syst Appl 38:2224–2234 31. Herlocker JL, Konstan J, Terveen L, Riedl J (2004) Evaluating
10. Berry MW, Dumais ST, O’Brien GW (1995) Using linear algebra collaborative filtering recommender systems. ACM Trans Inf
for intelligent information retrieval. SIAM Rev 37:573–595 Syst 22(1):5–53

123
Neural Comput & Applic (2012) 21:1017–1029 1029

32. Huang MJ, Chen MY, Lee SC (2007) Integrating data mining 53. Ryan PB, Bridge D (2006) Collaborative recommending using
with case-based reasoning for chronic diseases prognosis and formal concept analysis. Knowl Based Syst 19(5):309–315
diagnosis. Expert Syst Appl 32(3):856–867 54. Sarwar B (2001) Sparsity, scalability, and distribution in rec-
33. Janner T, Schroth C (2007) Web 2.0 and SOA: converging ommender systems, PhD thesis, University of Minnesota
concepts enabling the internet of services. IT Professional, pp 36– 55. Sarwar B, Karypis G, Konstan J, Riedl J (2000) Application of
41 dimensionality reduction in recommender systems—a case study.
34. Jenssen R, Erdogmus D, Hild KH, Principe JC, Eltoft T (2007) In: Proceedings of the KDD workshop on WebKDD
Information cut for clustering using a gradient descent approach. 56. Sarwar B, Karypis G, Konstan J, Riedl J (2002) Incremental
Pattern Recogn 4(3):796–806 singular value decomposition algorithms for highly scalable
35. Kröse B, Van Der Smagt P (1996) An introduction to neural recommender systems. In: Fifth International Conference on
networks, 8th edn. University of Amsterdam, Amsterdam computer and information technology (ICCIT 2002), pp 399–404
36. Lee KC, Kim JS, Chung NS, Kwon SJ (2002) Fuzzy cognitive 57. Sarwar D, Karypis G, Konstan J, Riedl J (2001) Item-based
map approach to web-mining inference amplification. Expert Syst collaborative filtering recommendation algorithms. In: Proceed-
Appl 22:197–211 ings of the 10th International World Wide Web Conference
37. Liu B (2007) Web data mining: exploring hyperlinks, contents (WWW’01), pp 285–295
and usage data. Springer, New York 58. Sevaux M, Dauze‘re-Pe0 re‘s S (2003) Genetic algorithms to
38. Maheswara Rao VVR, Valli Kumari V, Raju KVSVN (2010) A minimize the weighted number of late jobs on a single machine.
plausible comprehensive web intelligent system for investigation Eur J Oper Res 151:296–306
of web server behavior adaptable to incremental mining. Int J 59. Schafer JB, Konstan JA, Riedl J (2002) Meta-recommendation
Database Manag Syst (IJDMS) 2(3):95–114. doi : 10.5121/ijdms. systems: user-controlled integration of diverse recommendations.
2010.230995 In: Proceedings of the 11th international conference on information
39. Mithas S, Ramasubbu N, Krishnan MS, Fornell C (2007) and knowledge management (CIKM 2002), McLean, VA,
Designing web sites for customer loyalty across business pp 43–51
domains: a multilevel analysis. J Manage Inf Syst 23(3):97–127 60. Spiliopoulou M, Mobasher B, Berendt B, Nakagawa M (2003) A
40. Mobasher B (2006) Web usage mining. In: Wang J (ed) Ency- framework for the evaluation of session reconstruction heuristics
clopedia of data warehousing and mining, idea group in web usage analysis. INFORMS J Comput 15(2):171–190
41. Mobasher B (2007) Data mining for web personalization. In: 61. Srebroan N, Jaakkola DT (2003) Weighted low-rank approxi-
Brusilovsky P, Kobsa A, Neidl W (eds) The adaptive web: mations. In: Proceedings of the 20th international conference on
methods and strategies of web personalization. Lecture notes in machine learning, pp 720–727
computer science, vol 4321, Springer, Berlin 62. Srivastava J, Cooley R, Deshpande M, Tan P (2000) Web usage
42. Montoya-Torres JR, Gutierrez-Franco E, Mayorga CP (2010) mining: discovery and applications of usage patterns from web
Project scheduling with limited resources using a genetic algo- data. SIGKDD Explor 1(2):12–23
rithm. Int J Project Manage 28:619–628 63. Ting CK (2005) On the mean convergence time of multi-parent
43. Mulvenna MD, Anand SS, Buchner AG (2000) Personalization genetic algorithms without selection. In: Proceedings of the
on the net using web mining. Commun ACM 43(8):123–125 Eighth European conference on artificial life, Canterbury, UK,
44. Oliver A, Monmarché NG (2002) Venturini. Interactive design of pp 403–412
web sites with a Genetic algorithm. In: Proceedings of the IADIS 64. Theodoridis S, Koutroumbas K (2008) Pattern recognition, fourth
International Conference WWW/Internet, Lisbon, Portugal edition. Academic Press, London
45. Olson D, Shi Y (2007) Introduction to business data mining. 65. Tug E, Akiroglu MS, Arslan A (2006) Automatic discovery of the
McGraw-Hill Irwin, New York sequential accesses from web log data files via a genetic algo-
46. Ozel SA (2011) A web page classification system based on a rithm. Knowl Based Syst 19:180–186
genetic algorithm using tagged-terms as features. Expert Syst 66. Vozalis MG, Margaritis KG (2007) Using SVD and demographic
Appl 38:3407–3415 data for the enhancement of generalized collaborative filtering.
47. Perkowitz M, Etzioni O (2000) Adaptive web sites, association Inf Sci 177:3017–3037
for computing machinery. Commun ACM 43(8):152–159 67. Wu D, Yang Z, Liang L (2006) Using DEA-neural network
48. Pierrakos G, Paliouras G, Papatheodorou C, Spyropoulos C approach to evaluate branch efficiency of a large Canadian bank.
(2003) Web usage mining as a tool for personalization: a survey. Expert Syst Appl 31:108–115
User Model User-Adap Inter 13:311–372 68. Yeh IC, Lien CH, Ting TM, Liu CH (2009) Applications of web
49. Poon PW, Carter JN (1995) Genetic algorithm crossover operators mining for marketing of online bookstores. Expert Syst Appl
for ordering applications. J Comput Operat Res 22(1):135–148 36:11249–11256
50. Resnick P, Iacovou N, Suchak M, Bergstrom P, Riedl J (1994) 69. Zhang X, Edwards J, Harding J (2007) Personalized online sales
GroupLens: an open architecture for collaborative filtering of using web usage data mining. Comput Ind 58(8–9):772–782
netnews. In: Proceedings of 1994 ACM conference on computer 70. Zang H, Zhang S, Hapeshi K (2010) A review of nature-inspired
supported cooperative work. ACM Press, New York, pp 175–186 algorithms. J Bionic Eng 7:232–237
51. Resnick P, Varian HR (1997) Recommender systems. Commun
ACM 40(3):56–58
52. Rubin PA, Ragatz GL (1995) Scheduling in a sequence depen-
dent setup environment with genetic search. Comput Oper Res
22(1):85–99

123