Using Genetic Algorithms for Measuring the

Similarity Values between Users in Collaborative

Filtering Recommender Systems
Bushra Alhijawi
Hashemite University
Computer Information Systems
Zarqa, Jordan
Email: [email protected]
Yousef Kilani
Hashemite University
Computer Information Systems
Zarqa, Jordan
Email: [email protected]

Abstract—Recommender systems aim to help web users to find The basic principle of RS is the expectation that a group
only close information to their preferences rather than searching of users similar to one given user, (i.e. the group that have
through undifferentiated mass of information. Currently, col- rated an important number of elements in a similar way to the
laborative filtering is probably the most known and commonly user) can be used to adequately predict that group ratings on
used recommendation approach in recommender systems. In this products which the user has no knowledge of [2].
paper, we present a new genetic algorithms-based recommender
system, SimGen, that computes the similarity values between Bobadilla et al. [4] categorized RSs into three major groups:
users without using any of the well-known similarity metric collaborative filtering [3], [5], [6], content-based filtering [7]
calculation algorithms like Pearson correlation and vector cosine- and demographic filtering [8]. Collaborative filtering RSs
based similarity. The results obtained present 46% and 38% aggregate ratings or recommendations of objects (e.g. films,
improvements in prediction quality and performance, respectively books, videos, etc.), recognize commonalities between users on
when compared with other techniques.
the basis of their ratings and generate new recommendations
Keywords–collaborative filtering; recommender system; genetic based on inter-user comparisons [9]. Currently, collaborative
algorithms; similarity functions. filtering is probably the most known and commonly used
recommendation approach in RS [10]. In a content-based
I. I NTRODUCTION filtering systems, the objects of interest are defined by their
associated features [9] based on user choices made in the past
In everyday life, people rely on recommendations from other (e.g. in a web-based e-commerce RS, if the user purchased
people by spoken words, reference letters, news reports from some fiction films in the past, the RS will probably recommend
news media, general surveys, travel guides and so forth [1]. a recent fiction film that he has not yet purchased on this
Recommender systems (RSs) assist and augment this natural website) [4].
social process to help people sift through available books,
articles, web pages, movies, music, restaurants, jokes, grocery The content-based filtering RS makes recommendations by
products and so forth to find the most interesting and valuable analyzing the contents like text, images and sounds. In other
information for them [1]. Currently, the fast increase of Web words, content filtering tools select the right information for
2.0 has led to the proliferation of collaborative websites in the right people by comparing representations of the content
which the number of elements that can be recommended (e.g contained in the documents to the representations of the
blogs) can be increase significantly when introduced (and content that the user is interested in [11]. The main difference
not voted) by the users. Which generate new challenges for between CF and content-based RSs is that CF only uses the
researchers in the field of RSs [2]. RSs are currently being ap- user-item ratings data to make predictions and recommenda-
plied in many different domains [3] that can be found in many tions, while content-based RSs rely on the features of users and
modern applications such that DVD rental provider Netflix1 , items for predictions [12]. Demographic RSs aim to categorize
the online book retailer Amazon.com2 and the prediction of the user based on personal attributes (i.e. sex, age, county,
particular stock to the users in the PredictWallStreet company. etc) to make recommendations based on demographic classes
[9]. On the other hand, other RSs are hybrid systems that
1 www.netflix.com combine two or more groups of the previous three groups.
2 www.amazon.com Hybrid filtering is usually based on bio inspired or probabilistic
methods such as genetic algorithms (GA), fuzzy genetic, neural designed a recommender that exploits advantages of hybrid
networks, Bayesian networks, clustering and latent features [4]. CF for high quality prediction and recommendation. They pre-
The hybrid user-based fuzzy CF method [13] is an example of sented a GA-based approach for supporting combined modes
hybrid filtering algorithm. of CF to recommend new items for a particular user based on
Our interest in this paper is the employment of the genetic his previous likes or the opinions of other like-minded users.
algorithms in CF RSs which proposed by [14], [15], [16], Salehi et al. [18] proposed a hybrid RS for learning materials
[17]). In this paper, we present a novel genetic algorithms- based on their attributes. This system consists of two modules
based CF RS, SimGen3 . The fundamental objective of SimGen process. The weights of implicit or latent attributes of materials
is to improve the results, the accuracy and the speed of CF for learner are used as chromosomes in genetic algorithms.
RSs by developing a new similarity metric. SimGen computes Then it optimizes the weights according to the historical rating.
the similarity metric using the genetic algorithms and does Al-Shamri and Bharadwaj [19] developed a hybrid fuzzy-
not use any of the common similarity functions like Pearson genetic RS by employing GA to evolve appropriate weights
correlation and vector cosine-based similarity. Initially, in for each feature of the user model. In addition, they proposed
SimGen, the population consists of a set of members, each a novel fuzzy distance metric to match users and user model
represents a similarity metric. Where each member contains that enables hybrid filtering which reduces the system and
a random generated similarity value in the range of [0..1] computational time.
between every two users, then we split the data for next Jia et al. [20] designed a RS that uses the genetic algorithms.
process in two parts: training data and testing data. The training It selects weights and threshold values to be used in the
data is used in order to measure the fitness of the member similarity calculation. This system selects neighbors based on
when selecting the parents for mating after the crossover and the similarity which then gives the recommendation based on
mutation operators. The fitness and the testing data is used to trust. It considers the similarity and trust as necessary and
test the accuracy of the prediction. considers the characteristics of the users and the items own
The work presented in this paper makes the following characteristics to improve the effect. Finally, the system gets
contribution: the more accurate selection of the best neighbors to make a
further improvement of the recommendation accuracy.
• A new mechanism for calculating the similarity values
All the reviewed papers use a similarity metric in order to
between users using GA is proposed.
calculate the similarity between users or items. This similarity
The rest of this paper is organized as follows. Section II metric is either one of the well-known matrices or it is
covers the related work in the RS and GAs. Section III shows based on calculating the statics or historical data about the
the proposed algorithm, SimGen, and the genetic operators rating of the users. Our algorithm, SimGen, does not use
that are used. Section IV clarifies the evaluation method that any similarity function. It starts by giving a random initial
is used and compares SimGen with other techniques. Section similarity value between every two users and this similarity
V concludes the paper and presents the future work. value is enhanced from generation to generation as SimGen
runs. This enhancement is done by the direction of the training
data, and SimGen does not require any additional information
II. R ELATED W ORK
provided by hybrid model like Bobadilla et al. [2], [20].
Bobadilla et al. [4] presented a recent survey of the RSs.
Su and Khoshgoftaar [1] introduced a comprehensive survey
III. U SING GA TO C OMPUTE THE M ETRIC
of collaborative filtering techniques. In this section, we review
the related work on the RS using GAs. The main objective is to improve the results of CF by
Bobadilla et al. [2] developed a genetic algorithms-based using GA to compute the similarity metric. SimGen does
RS, GARS, that is based on CF which use the rating of not use any of the common similarity functions like Pearson
users to compute the similarity. Bobadilla et al. [2] built a correlation and vector cosine-based similarity. It generates an
new similarity function that measures the similarity between initial random similarity value between every two users then
every two users. GARS makes statistics about every two users. it runs the GA. The proposed algorithm uses training data to
It calculates how many times both users rate items exactly compute the fitness in every generation and the testing data to
with a difference of 1, 2, 3, and 4. The later calculated value test the resulting similarity to compute the errors in prediction.
represents the difference between the rating of two users.
GARS associate a weight with each value then it computes
A. Initial Population
the weight by using the GAs.
Gao and Li [16] proposed a hybrid model to integrate The initial population consists of a set of individuals. We
outputs produced by every RS at the basis of GAs by con- represent each individual by a 2-dimensional array as shown
structing weight vectors that represent different forecasting in fig. 1. Initially, we generate a random similarity value in the
performances of each RS. The hybrid problem is translated range [0..1] between every two users for each individual. It is
into optimizing problem about weight vectors. Fong et al. [17] a common knowledge that the similarity values in RS ranges
between [-1..1] and the similarity value between the user and
3 The source code can be obtained by emailing one of the authors itself is 1.
 TABLE I. S IM G EN COMPONENTS
The initial population consists of 50 individuals, each one contains
Initialization Initial population
a random similarity value in range [0..1] between every two users.
The Mean Absolute Error (MAE) is computed depending on the
Assessment Fitness Function
predicted rates of items that are rated by each of the training users.
Selection Roulette-Wheel selection (RW) technique is used.
Genetic Operators Crossover Uniform crossover operator is used with 50% probability.
Mutation Single point mutation technique with 20% probability.

User
U P Iu
1 2 3 ... x
1 1 1 X i=1 |piu − rui |
User 2 sim(2,1) 1 M AE = (2)
3
...
sim(3,1)
sim(...,1)
sim(3,2)
sim(...,2)
1
sim(...,3) 1
#U u=1 #Iu
x sim(x,1) sim(x,2) sim(x,3) sim(x,...) 1
Where, #U represents the number of training users and #Iu
Figure 1. similarity matrix between users represent the number of training items rated by the user u.
Example 1. Suppose that there are 5 users, 10 items, and each
B. Fitness Function of the 5 users rated at least three items as shown in fig. 2 and
there are two members m and n as shown in fig. 3. It is clear
The fitness function is a major component of GA that is that m is better than n since the accuracy of prediction resulted
used to evaluate the individuals in the population. In SimGen, from m is higher than resulted from n as shown in fig. 4. The
the role of the fitness function is to measure the optimality of MAE (equation 2) result from using the similarity matrix m
the individual (i.e. similarity array). is 0.958 while it is 1.3 when use the similarity matrix n. ⋄
In order to calculate the fitness function for each individual,
we use the similarity metric (i.e the individual) to compute I1 I2 I3 I4 I5 I6 I7 I8 I9 I10
U1 3 . 5 1 1 . 2 . . .
the prediction of each user rated items for each training user. U2 1 2 . 3 . . . . . .
The prediction of user x on item i is given by the following U3
U4
.
. 2
. 4
. 2
. .
.
1
. 5
. 3
3
5
.
.
4
well-known formula: U5 4 . 4 2 2 3 . . 3 .

P kx i
Figure 2. User-Item rating matrix
n=1 [sim(x, n) ∗ (rn − r̄n )]
pix = r̄x + P kx (1)
n=1 sim(x, n) Member M Member N
U1 U2 U3 U4 U5 U1 U2 U3 U4 U5
Where, U1 1 U1 1
U2 0 1 U2 0.4 1
• r̄x represents the average ratings made by user x for the U3 0.3 0 1 U3 0 0.3 1
training items x rated. U4 0.11 0.7 0.2 1 U4 0 0 0.5 1
U5 0.71 0.2 0.8 0.2 1 U5 0.2 0.5 0.3 0.2 1
• sim(x, n) is the similarity value between users x and n.
This value is taken from the individual. Figure 3. Members M and N
• rni represents the rating of user n on item i.
• kx is the set of the neighbor users to user x. In SimGen, Prediction Rates M
I1 I2 I3 I4 I5 I6 I7 I8 I9 I10
we consider kx to be the set of the training users (80% U1 3.4 . 3.4 1.3 1.4 . 1.2 . . .
of the all the users) which x rated. U2
U3
3
.
1.8
.
.
3.6
1.8
.
.
.
.
3.25
.
.
.
3.1
.
3.25
.
.
U4 . 2.2 . 2.8 . . 2.8 2.9 . 3.2
The MAE (equation 2) is obtained by computing the differ- U5 3 . 4.6 1.9 1.6 0.75 . . 4.75 .

ence of the actual ratings with the predicted ratings that are
Prediction Rates N
produced depending on the similarity array. We calculate the I1 I2 I3 I4 I5 I6 I7 I8 I9 I10
predictions for all the training users on all the training items U1
U2
1.4
3.8
.
2.4
3.4
.
2.7
1.8
1.4
.
.
.
2.4
.
.
.
.
.
.
.
based on the similarity array in order to compute the MAE U3 . . 4.25 . . 3.25 . 3.45 3.25 .
U4 . 3.2 . 2.2 . . 3.2 2.95 . 3.2
of the RS. The lowest fitness value for individual z means U5 2.4 . 3.3 2.4 1.6 0.75 . . 4.75 .
that individual z provides closest predictions. The objective of
running SimGen is to obtain an individual with the smallest Figure 4. The prediction rates of members M and N
MAE. Therefore, the fitness value for a member m is better
than the fitness value for a member n means that m is a better
similarity metric than n since the users have predicted values C. Genetic Operators
according to the similarity metric in m closer to their actual
rate than those in n. Consider the following example which SimGen uses the selection, crossover and mutation opera-
illustrates the idea. tors. The crossover and mutation probability is 0.5 and 0.2
respectively. We select the individuals for mating based on testing data (i.e. 20%). All the algorithms were built using
their fitness value. As a selection technique, the roulette-wheel Visual Basic programming language and the CPU time is
parent selection algorithm(RW) is used to select the parents measured using Process Explorer software. We used a PC of
for mating. RW gives high probability value for the good 2 GB RAM with Dual CPU 2.16 GHz.
individuals and low probability value for the bad individuals TABLE II. M AIN PARAMETERS USED IN THE EXPERIMENTS
to be selected. Note that both the good and the bad individuals
Dataset Precision/Recall K-Neighbors Test user % Test item % Genetic runs
have chances to be selected, however the good individuals have N
higher chances. The uniform crossover operators is applied Movielens
Synthetic data
{2,...,20}
{2,...,20}
{20,40,60,...,200}
{2,4,6,...,20}
20%
20%
20%
20%
100
100
and two new children are generated after each round of
crossovers. These children replace the worst two members
The results of SimGen are compared with the ones obtained
in the current population in case they have better fitness.
from the other approaches using traditional metric on CF
Otherwise, they are discarded. We apply the mutation operation
RS: Mean Absolute Error (MAE), precision, and recall. Each
on the children produced as a result of the crossover operation
of the 4 techniques: SimGen, Bobadilla et al. [2], Pearson
with a probability of 0.2.
correlation and cosine metric were ran for 100 times and the
average of the 100 runs is taken.
IV. E XPERIMENT Table III present the average CPU time in seconds was taken
In this section, we experimentally prove that the prediction to make the prediction by the 4 algorithms. It is clear from
quality and the performance of our algorithm, SimGen, is the results that that there is a high improvement in speed
better than the results obtained by other techniques like cosine using SimGen. SimGen takes 15.8 while GARS [2] takes
metric, Pearson correlation, and genetic-based algorithms (i.e. 22.65 seconds, cosine metric takes 27.6 seconds, and Pearson
GARS [2]). We have used in the experiments two different correlation takes 25.9 seconds. This means that SimGen is
datasets: one and half times (22.65/15.8 = 1.5, 25.9/15.8 = 1.6) faster
than GARS [2] and Pearson correlation, and nearly two times
• Movielens. Movielens 4 database. The Movielens (27.6/15.8=1.7) faster than cosine metric.
presents a common reference in RS research that consists
of 100,000 ratings (1-5) from 943 users on 1682 movies. TABLE III. T HE PERCENTAGE OF IMPROVEMENT IN SPEED RESULTING
FROM S IM G EN
Each user has rated at least 20 movies.
SimGen GARS [2] Cosine metric Pearson correlation
• Synthetic data. Which is a random biased data as Mar- CPU time 15.8s 22.65s 27.6s 25.9s
manis and Babenko [21] did to test their recommenda- Improvement =
algorithm(A)−SimGen 30.2% 42.79% 39%
tion engine which they called Delphi. This data consists algorithm(A)

of 100 users, u1 , u2 ,. . ., u100 , 300 movies, i1 , i2 ,. . .,

i300 . The users and items are splitted into sets as follows: Figs. 5, 6 and 7 show the results of the MAE, recall,
and precision measures of the 4 techniques. The number of
◦ U40 = {u1 ,. . ., u40 }, contains 40% of the total
neighbor users for a user is represented by the constant k in
users.
fig. 5. These neighbors vary between 20 to 200 and 2 to 20
◦ U60 = {u41 ,. . ., u100 }, contains the other 60% of
when using Movielens and synthetic data respectively.
the users.
◦ I40 = {u1 ,. . ., u40 }, contains 40% of the total The results of recall and precision metrics show in figs. 6
items. and 7 respectively. Fig. 6 presents that the quality is improved
◦ I60 = {u41 ,. . ., u100 }, contains the other 60% of in any number of recommendation. Also in fig. 7, it is clear
the items. that the quality measure is improved for any value used in the
Three constraints were placed to build this data: number of recommendation using both datasets.
1) Each user has rated n movies, where n is a
random number between 10 and 20. V. C ONCLUSION
2) n items from I40 assigned randomly to each user In this paper, we presented a new genetic-based algorithm,
from U40 , a random rate set to 4 or 5. SimGen, for computing the similarity metric. We showed that
3) n items from I60 assigned randomly to each user SimGen does not use any of the existing similarity functions
from U60 , a random rate set to 1, 2 or 3. like Pearson correlation and vector cosine-based similarity.
The synthetic random data built in a such a way, can be In addition, it does not require the additional information
easily determined if the result of running any algorithm provided by the hybrid models. Initially, SimGen gives the
using it is correct answer. For instance, the prediction of similarity between every two users a random value. Then, it
any user from U60 for any item of I60 should be near runs to adjust this value based on the training data.
(1, 2 or 3). Otherwise, the prediction is not accurate. In order to evaluate SimGen, the traditional metric on CF
Table II shows the main parameter used in the experiments. RS: MAE, precision, and recall. We proved that SimGen
Each dataset was divided into training data (i.e. 80%) and achieved high performance in accuracy and quality compared
to the current state-of-the-art genetic-based algorithms like
4 downloadable from http://grouplens.org/datasets/movielens/ GARS [2], cosine metric, and Pearson correlation. In the
 A B

0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6

MAE

MAE
0.5 0.5
0.4 0.4
0.3 SimGen 0.3 SimGen
Bobadilla et al. [2] Bobadilla et al. [2]
0.2 0.2
Pearson correlation Pearson correlation
0.1 Cos 0.1 Cos
0 0
0 20 40 60 80 100 120 140 160 180 200 0 2 4 6 8 10 12 14 16 18 20
K-neighbors K-neighbors

Figure 5. The MAE measure of the Bobadilla et al. [2], the Pearson correlation, the cosine metric, and the SimGen algorithms. A) Movielens. B)Synthetic
data.

A B

SimGen 0.9 SimGen

Bobadilla et al. [2] Bobadilla et al. [2]
Pearson correlation 0.8 Pearson correlation
0.9
Cos 0.7 Cos
0.8
0.7 0.6
Recall

Recall
0.6 0.5
0.5 0.4
0.4
0.3
0.3
0.2
0.2
0.1 0.1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
Number of Recommendations Number of Recommendations

Figure 6. The recall quality measure of the Bobadilla et al. [2], the Pearson correlation, the cosine metric, and the SimGen algorithms. A) Movielens.
B)Synthetic data

compression process, two data sets were used: synthetic data R EFERENCES
and Movielens. The results obtained present that SimGen is:
[1] X. Su and T. Khoshgoftaar, “A Survey of Collaborative Filtering
Techniques,” Journal of Advances in Artificial Intelligence, 2009.
• Using synthetic data: [2] J. Bobadilla, A. Ortega, F.and Hernando, and J. Alcalá, “Improving Col-
1) One and half times faster than GARS [2] and laborative Filtering Fecommender System Results and Performance us-
ing genetic algorithms,” Journal of Knowledge-Based Systems, vol. 24,
Pearson correlation. 2011, pp. 1310–1316.
2) Two times faster than cosine metric. [3] P. Moradi and S. Ahmadian, “A Reliability-based Recommendation
3) 33% improvement in prediction quality than Method to Improve Trust-Aware Recommender Systems,” Journal of
GARS [2]. Expert Systems with Applications, vol. 42, no. 21, 2015, pp. 109–132.
4) 69.5% improvement in prediction quality than [4] J. Bobadilla, F. Ortega, A. Hernando, and A. Gutiérrez, “Recommender
cosine metric. Systems Survey,” Journal of Knowledge-Based Systems, vol. 46, no. 8,
5) 55.5% improvement in prediction quality than 2013, pp. 109–132.
Pearson correlation. [5] S. Ghazarian and M. Nematbakhsh, “Enhancing Memory-based Collab-
orative Filtering for Group Recommender Systems,” Journal of Expert
Systems with Applications, vol. 42, no. 7, 2015, pp. 3801–3812.
• Using Movielens data: [6] K. Christidis and G. Mentzas, “A Topic-based Recommender System
for Electronic Marketplace Platforms,” Journal of Expert Systems with
1) One and half times faster than GARS [2] and Applications, vol. 40, no. 11, 2013, pp. 4370–4379.
Pearson correlation. [7] G. Adomavicius and A. Tuzhilin, “Context-Aware Recommender Sys-
2) Two times faster than cosine metric. tems,” Recommender Systems Handbook, F. Ricci, et al. (Ed.), 2011,
pp. 217–253.
3) 14.66% improvement in prediction quality than
GARS [2]. [8] I. Ma, H.and King and M.-R. Lyu, “Learning to Recommend with
Explicit and Implicit Social Relations,” ACM Transactions on Intelligent
4) 70% improvement in prediction quality than co- Systems and Technology, vol. 2, no. 3, 2011.
sine metric. [9] R. Burke, “Hybrid Recommender Systems: Survey and Experiments,”
5) 46.5% improvement in prediction quality than Journal of User Modeling and User-Adapted Interaction, 2002, pp. 331–
Pearson correlation. 370.
 A B

1.1 SimGen 1.1 SimGen

1 Bobadilla et al. [2] 1 Bobadilla et al. [2]
0.9 Pearson correlation 0.9 Pearson corelation
0.8 Cos 0.8 Cos

Precision

Precision
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20
Number of Recommendations Number of Recommendations

Figure 7. The precision quality measure of the Bobadilla et al. [2], the Pearson corelation, the cosine metric, and the SimGen algorithms. A) Movielens.
B)Synthetic data

[10] Q. Shambour and J. Lu, “A Hybrid multi-criteria semantic-enhanced

collaborative filtering approach for personalized recommendations,”
IEEE WICACM International Conference on Web Intelligence, 2001.
[11] J. Herlocker, J. Konstan, A. Borchers, and J. Riedl, “An Algorithmic
Framework for Performing Collaborative Filtering,” In: 22nd annual
international ACM SIGIR conference on Research and development in
information retrieval, 1999, pp. 230–237.
[12] L. Si and R. Jin, “Flexible Mixture Model for Collaborative Filtering,”
In: the 20th International Conference on Machine Learning, vol. 2,
2003, pp. 704–711.
[13] L. Son, “HU-FCF: A Hybrid User-based Fuzzy Collaborative Filtering
Method in Recommender Systems,” Journal of Expert Systems with
Applications, vol. 41, no. 15, 2014, pp. 6861–6870.
[14] K. Kim and H. Ahn, “A Recommender System Using GA K-means
Clustering in an Online Shopping Market,” Journal of Expert Systems
with Applications, vol. 34, no. 2, 2008, pp. 1200–1209.
[15] ——, “Using a Clustering Genetic Algorithm to Support Customer
Segmentation for Personalized Recommender Systems,” In: 13th Inter-
national Conference on AI, Simulation, and Planning in High Autonomy
Systems, 2005, pp. 409–415.
[16] L. Gao and C. Li, “Hybrid Personalized Recommended Model Based
on Genetic Algorithm,” In: 4th International Conference on Wireless
Communications, Networking and Mobile Computing (WiCOM), 2008,
pp. 1–4.
[17] S. Fong, Y. Ho, and Y. Hang, “Using Genetic Algorithm for Hybrid
Modes of Collaborative Filtering in Online Recommenders,” In: 8th
International Conference on Hybrid Intelligent Systems (HIS), 2008,
pp. 174–179.
[18] M. Salehi, M. Pourzaferani, and S. Razavi, “Hybrid Attribute-based
Recommender System for Learning Material using Genetic Algorithm
and a Multidimensional Information Model,” Egyptian Informatics
Journal, vol. 24, no. 1, 2013, pp. 67–78.
[19] M. Al-Shamri and K. Bharadwaj, “Fuzzy-Genetic Approach to Rec-
ommender Systems based on a Novel Hybrid User Model,” Journal of
Expert Systems with Applications, vol. 35, no. 3, 2008, pp. 1386–1399.
[20] Y. Jia, Q. Ding, D. Liu, J. Zhang, and Y. Zhang, “Collaborative Filtering
Recommendation Technology based on Genetic Algorithm,” Journal of
Applied Mechanics and Materials, vol. 599–601, no. 3, 2014, pp. 1446–
1452.
[21] H. Marmanis and D. Babenko, Algorithms of the intelligent web
minning, 2nd ed., 2009.