Mapping Evaluation for Semantic Browsing

Abstract

The paper contributes to the problem solving in semantic browsing and analysis of scientific articles. With reference to presented visual interface, four – the most popular methods of mapping including own approach - MDS with spherical topology, have been compared. For a comparison quantitative measures were applied which allowed to select the most appropriate mapping way with an accurate reflection of the dynamics of data. For the quantitative analysis the authors used machine learning and pattern recognition algorithms and described: clusterization degree, fractal dimension and lacunarity. Local density differences, clusterization, homogeneity, and gappiness were measured to show the most acceptable layout for an analysis, perception and exploration processes. Visual interface for analysis how computer science evolved through the two last decades is presented on website. Results of both quantitative and qualitative analysis have revealed good convergence.

Full text

Abstract—The paper contributes to the problem solving in
semantic browsing and analysis of scientific articles. With
reference to presented visual interface, four – the most popular
methods of mapping including own approach - MDS with
spherical topology, have been compared. For a comparison
quantitative measures were applied which allowed to select the
most appropriate mapping way with an accurate reflection of the
dynamics of data. For the quantitative analysis the authors used
machine learning and pattern recognition algorithms and
described: clusterization degree, fractal dimension and
lacunarity. Local density differences, clusterization,
homogeneity, and gappiness were measured to show the most
acceptable layout for an analysis, perception and exploration
processes. Visual interface for analysis how computer science
evolved through the two last decades is presented on website.
Results of both quantitative and qualitative analysis have
revealed good convergence.
I. INTRODUCTION
onlinear growth of scientific writing imposes a new
forms of academic databases management. The latter
includes the both retrieval and analytical exploration. Ana-
lysts, science of science professionals, science policy makers
need various computing, statistical and visualization tools to
monitor how science or scientific domains evolved.
N
Authors designed and described in a series of papers [1],
[2] the visual interface for analysis of dynamics of computer
science through the two last decades1. Screenshot of spheri-
cal application is shown on Figure 1. Users can interact, ma-
nipulate and browse the data and see how graphical pattern
change in time. The nodes represent scientific articles from
digital library and the colour – appropriate thematic cate-
gory. Similarity metrics was based on semantic relations be-
tween documents [2]. In order to generate 3D layout, multi-
dimensional scaling (MDS) technique was applied and en-
riched by Morse potential [1].
Overlapping spots show where the categories mutually in-
tegrate that means an articles at that location are semantically
similar. Visualization of classified documents reveals both
organization of digital library content as well as allows users
to track how it changes over time. This paper presents fur-
1http://www-users.mat.umk. pl/~garfi/vis2009v3/
ther study on visualization maps. Authors decided to test this
prototype regarding to mapping algorithms. Four mapping
methods were compared in terms of dynamics and analytical
possibility of output visualization. The next chapter shows
the outline of VxOrd, MDS, VOS, SOM as the most popular
methods.
II.MAPPING PREVIEW
How we, as analysts, perceive and understand the connec-
tions between data, depends on graphical layout. Thus, the
final structure of visualized knowledge can be drawn either
by spatial arrangement (2D or 3D) of analysis units or by the
relationship between nodes in graph or combination of these
two.
One of the basic ordination algorithm - VxOrd extends a
traditional force-directed approach [3]. VxOrd determines
the both number and size of clusters automatically based on
the data. Popular software for data mapping and visualiza-
tion - Gephi2 uses this technique. Due to Gephi users can an-
alyse large networks consisting of even millions of nodes.
The most popular technique for dimension reduction is
MDS, which involves minimizing the difference between
Euclidean and graph-theoretic distances. MDS has been
widely applied for constructing knowledge maps of authors,
articles, journals, and keywords [3]-[5]. The same satisfac-
tory representation of knowledge can be produced by use of
new mapping technique VOS introduced in series of works
by Van Eck and Waltman [6],[7]. The idea of VOS is to
minimize a weighted sum of the squared distances between
all pairs of items.
Dimension reduction can be also achieved in self orga-
nized maps – the kind of unsupervised neural network which
aimed to project high-dimensional data into a lower-dimen-
sional space [4]. The nodes (input vectors) form two dimen-
sional regular grid; node’s neighbourhood is defined to be all
connected nodes. During training process similar input vec-
tors stimulate adjacent neurons and therefore output SOM
map shows semantic relationships between data, where simi-
lar items are mapped close together. Comparing MDS and
VOS,
2http://gephi.github.io/
Veslava Osinska
Institute of Information Science
and Book Studies, Nicolaus
Copernicus University,
ul. Bojarskiego 1, 87-100 Toruń,
Poland
Email: wieo@umk.pl
Adam Jozwik
Institute of Biocybernetics and
Biomedical Engineering, Polish
Academy of Science, Warsaw;
Institute of Computer Science,
College of Social and Media
Culture, Toruń, Poland
Email: adamj346@wp.pl
Grzegorz Osinski
Institute of Computer Science,
College of Social and Media
Culture, ul. Starotoruńska 3
87-100 Toruń, Poland
Email: gos@fizyka.umk.pl
Mapping Evaluation for Semantic Browsing
Preprints of the Federated Conference on
Computer Science and Information Systems pp. 341–347
c
2015 341

Fig. 1 Interface screenshot - the prototype
of application for semantic browsing
researchers concluded that maps constructed using VOS ap-
proach provide a more satisfactory representation of the un-
derlying dataset [8]. MDS-VOS tests revealed the first one is
distance preserving while the second - topology preserving
technique [9],[10].
III. METHODOLOGY
The complex structure of large graphs commonly is mea-
sured by modularity. According to Fortunato [11], modular-
ity can be defined as the function which evaluates the good-
ness of partitions of graph and is defined by ties between
vertices, vertices and hubs. Structure described by linked ob-
jects does not match with character of present data. Authors
analyse the visualization of classified documents. Primary
categories are taken from library classification. All relation-
ships between collected classes and documents were prede-
fined by specialists permanently working on computer sci-
ence taxonomy. Visualized articles are assumed to reveal se-
mantics while keeping their thematic similarity. These new
correlations (down-top) between data allow to rich indepen-
dence of the original organization of items (top-down). As
every rigid scheme, it is characterized by adequacy and dis-
jointedness of subclasses. Therefore, evaluation metrics can
be based on spatial configuration of visual layout instead of
links outline. To analyze the final nodes distribution, image
processing methods were applied. Then evaluation of graphi-
cal pattern should be carried out taking into account the ac -
curacy, topology (space filling and capacity) and perception
abilities of users.
The authors present alternative approach: a sphere surface
has been selected as a target mapping space. There are some
arguments for a sphere surface: it “has no edges and there-
fore it is possible to represent not only local similarities but
also large-scale ones regarding the whole space. The benefit
of a curved surface in comparison to a plane one is a more
capacious exploration space” [1,10]. 3D visualization is a
popular but also challenging method in large dataset map-
ping and modelling.
A. Assumptions
By implication, evaluation process will touch how to fit
interface to the requirements of analysts and domain experts.
The study is based on the following assumptions.
1. A given visualization layout might serve as a graph-
ical interface for the exploration and semantic re-
trieval of scientific articles. From this point of view
the most important is configuration on the bottom
level (documents) – then can be evaluated the spa-
tial distribution of nodes.
2. Current modifications by editors of the original
classification are aiming at its improvement. The
classification reflects the most of current changes in
computer science. Quickly developing categories
will form dense clusters and overlap each other.
These tendencies must be visible on visualization
maps generated for different time periods.
3. In the construction of the ergonomic user interface,
such features as capacity, homogeneous distribution
and edgelessness must be taken into consideration.
Short movie3 shows how three dimensional configuration al-
lows the user to analyze semantic distribution of articles and
its behavior in time.
B. Evaluation steps
On Figure 2 we can see the elements of evaluation
process. Continuity characteristics is crucial for present
study. For this purposes clusterization potential was vali-
dated by machine learning and image recognition algorithms.
Structural complexity can be evaluated by fractals analysis.
Quantitative measures are different for several maps and the
changes tendency are essential too.
All dimension reduction methods determine the arrange-
ment of classes and subclasses nodes. Documents distribu-
tion was calculated by using geometrical rules in 2D or 3D
space [10]. Obtained pattern became the basic material for
comparison and further study.
If we plan to involve users to scientific domain analysis,
visualization interface must be user-friendly and carry good
navigational features. Another usefulness of such application
is retrieval of semantically similar documents. Precision in
this case will be an appropriate measure of this visual search-
ing system.
C. Research material
Visualization maps were obtained by using the same data
but distinct in terms of data configuration (like matrixes ver-
sus data pairs), mapping algorithms and space topology. The
series of every ten-year layouts show the changes of pattern
and thus the evolution of the ACM classification and com-
puter science knowledge (see Appendix). An insight into the
differences in graphical patterns could reveal the most and
the least complex structures due to human perception. The
system of human perception is able to recognize a natural tex
3www.wizulizacjainformacji.pl/unas/interface.avi
342 PREPRINTS OF THE FEDCSIS. Ł ´
OD ´
Z, 2015

Fig. 2 Decision map of visualization evaluation steps
ture appearing in nature as a result of evolutionary adapta-
tion. The human vision allows determining approximately
whether the perceived structure differs one from another in
terms of complexity[12].
For example, the first series of maps is characterized by a
relative even distribution while VxOrd by data grouping on
edges and bends. Furthermore, VxOrd and VOS distributions
are highly limited to the output geometry [10]. The result
may be a non-effective space for navigation in those cases.
But human perception cannot be one of the main criteria for
comparison and estimation of visual layouts, although useful
in the final conclusions. Quantitative approach requires that
the authors analyze local density differences and quantify
clusterization, rarefaction, homogeneity and porosity.
Output maps can be described by both density and colour
of nodes. If information about the main thematic category as-
signment is excluded (i.e. the colour), the clusters can be
identified by density only. Consequently, clusterization and
its changes can deliver information on how knowledge ad-
vances and how knowledge organization changes throughout
two decades, independently of the primary (original) classifi-
cation.
IV. QUANTITATIVE ANALYSIS OF MAPS
A. Clusterization and its dynamics
To identify clusters on given maps we used the most pop-
ular partition technique, based on distances between points
and/or points to centroids – k-means clustering [11]. Algo-
rithm aims to minimize the within-cluster sum of squares:
∑
i=1
k∑
xj∈Si
‖xj−ci‖2
(1)
where Si indicates the subset of points in the i-cluster, ci – the
centroid of cluster.
The disadvantage of this algorithm in our assignment is
the requirement to know the number of clusters. To find
optimal number of classes we modeled the data by a set of
Gaussian distributions [13].
The number of components can be estimated by the
Bayesian Information Criterion [14] (BIC), which is based
on a penalization of the observed log-likelihood – the func-
tion of x, θ. The preffered model is the one with the lowest
value of BIC which decides about the number of clusters.
Thus the 6 clusters are recognized for the MDS –sphere (au-
thors method) and SOM and 8 for both VxOrd and VOS.
These new clusters reorqanize initial data assignment to the
11 main initial categories, coded by colour. By k-means clus-
tering the centroids of clusters are found, demonstrated on
Figure 3 according to the the MDS –sphere map.
Dynamical characteristics of clustering are crucial for a fi-
nal evaluation of the presented approaches in terms of struc-
tural analysis. In any sequence of maps it is possible to find
the one with a highly developed clusterization just intu-
itively.
To evaluate clustering and its dynamics, a misclassifica-
tion rate offered by the standard k nearest neighbour (k-NN)
rule was used as a criterion. That error rate was estimated by
the leave-one-out method [15],[16]. The k-NN rule assigns
the classified object to the class most heavily represented
from among its nearest objects in the training set (i.e. nearest
neighbours). The reference set, also called a training set, is a
set of objects with a known class membership and in a cer-
tain sense it defines the considered classes. The leave-one-
out method consists in the classification of each object from
the reference set by the decision rule obtained from the train-
ing set decreased by the currently classified object. The ratio
of the number of misclassified objects to a numerical force
of the reference set estimates the above mentioned error rate
used as the clustering quality criterion.
The low error rate value denotes that the considered
classes (or clusters) differentiate easily, but high values of
the error rate mean that the classes overlap. The leave one
out method is very convenient in the case of classifiers based
on the k-NN rule since no training is required. This property
of the k-NN classifiers was intesively used for creating a
fuzzy k-NN rule proposed by one of the authors of the
present work [17] and for introducing the more sophisticated
pair-wise k-NN classifier [18].
Fig. 3 Data distribution with six clusters centroids
(along “horseshoe” shape).
VESLAVA OSINSKA, ADAM JOZWIK, GRZEGORZ OSINSKI: MAPPING EVALUATION FOR SEMANTIC BROWSING 343

In every series of visualizations it is possible to point at
the map with the clearest clustering structure – model map.
The other, as it can be assumed, develops towards clustering
pattern. If the algorithm for pattern (points) clustering of
model map is trained, other ones can be tested by using the
nearest neighbour method. Which one serves as a training set
and which one as a testing set can be found first by evalua-
tion of the standard k-NN classifier. The lower error rate
means a better clustering structure (bold numbers in Train-
ing part of Table 1).
The results of training allow selecting an apropriate
dataset (bold) for testing. During the testing phase (Testing
part) the clusterization quality can be tracked on the basis of
error rate changes. It is worth noting that comparison must
be made along rows, not columns, because of the use of dif-
ferent methods to generate patterns.
The data in Table 1 show that the clusterization increases
in one case – the first row, which characterizes the authors’
approach. It proves continuous changes of CCS towards
overlapping categories and reorganization needs.
B. Even distribution, FD and Lacunarity
The authors’ study of visualization maps relates to visual
interface: which method of mapping can deliver the best way
to explore a complex dataset of scientific articles?
Restrained homogenous distribution on a sphere surface
can be estimated by the volume of empty places. The more
holes in the pattern, the more heterogeneous it is. The appro-
priate parameter is Lacunarity - the degree of holes distribu-
tion having the lowest value for indeterminate structure. La-
cunarity is often used in medical imaging for detection of
structural changes in bone texture on radiographs [19].
Visualization maps can be considered as flat textures asso-
ciated with the patterns of documents nodes distribution. La-
cunarity λ is defined as:
λϵ
,g =CV
ϵ
,g
2=
(
σ
μ
)
ϵ
, g
2
(2)
where σ is the standard deviation and μ is the mean for
pixel per box at this size ε, in a box at this orientation g.
Lacunarity pertains to both gaps and heterogeneity. To
simplify, the more gappiness in the image (i.e. sparsely
occupied maps), the higher lacunarity. Some recent research
has shown that there is a correlation between lacunarity and
fractal dimension, FD [20],[21].
The FD is a complexity indicator with a non-integer value.
The fractal dimension could be characterized as a scale of
transition to homogeneity and is therefore very practical in
the dynamics study case. Because the maps were generated
by three different mapping algorithms they present distinct
homogeneity, what is according to our assumption, one of
the criteria of the ergonomic visual interface.
The values of lacunarity and the FD for every map are
shown in Table 2. The highest value in each row (bold
numbers in the first part of Table 2) indicates a map with the
large porosity (gappiness). Bold FD values in the second part
of the same table means the best formed structure (implicitly
clear clusterization). The first row data presents a continuous
growth of complexity degree with simultaneously dense
occupation (high FD and lowest lacunarity values). Other
(VOS, SOM and VxOrd) demonstrate oscillations are
difficult to interpret. Consequently, the dynamics of each
index across time for every method was evaluated.
Lacunarities of every method should not be compared
because of different spanning geometry.
If should be taken into consideration that the fractal
dimension for random (or pseudo-random) distribution
equals 2.77, the more does the FD tend to this value, the
pattern is more homogeneous [22]. And inversely, the low
FD (bold numbers in Table 2) means the distribution
resembles linear. A stable structural change (in contrary of
step change) in time is proved by the authors’ method.
V. DISCUSSION
According to authors’ conception, in order to measure dy-
namics of graphical patterns we need to focus on how com-
plexity evolves. Therefore clustering resolution has been
tested by use of machine learning and pattern recognition al-
gorithms.
TABLE I.EVALUATION OF CLUSTERIZATION BY K-NEAREST NEIGBOURS METHOD FOR EACH VISUALIZATION MAP BASED ON ERROR RATES.
Phase Training Testing
1988 1998 2009 1988 1998 2009
Authors’ method 0.0140 0.0115 0.0074 0.8890 0.8300 -
VOS 0.0119 0.0078 0.0110 0.9011 -0.8710
VxOrd 0.0084 0.0107 0.0045 0.8707 0.7898 -
SOM 0.0143 0.0121 0.0128 0.8480 0.8090
344 PREPRINTS OF THE FEDCSIS. Ł ´
OD ´
Z, 2015

The authors proposed the qualitative measures for evalua-
tion of structural changes of pattern: FD and lacunarity.
In general, we can conclude: the larger complexity degree,
the lower randomness. On the other hand, the complex struc-
ture is also can be determined by the clustering level. The
current findings confirmed by presented measures (Table 1,
Table 2) show that clusters have become more explicit with
time at the maps generated by authors’ approach and at the
same time tend to uniform distribution (lowest lacunarity and
FD resembles the value of random distribution). The high la-
cunarity informs about dense network of holes among others,
due to overlapping pattern. Next acceptable technique in the
terms of changes continuity is SOM. Qualitative approach to
compare visualization maps [10] shows the similar results:
both MDS-sphere and SOM reveal consequence in dynamics
changes, moreover VOS and VxOrd – inappropriate topol-
ogy for data exploration [10, Appendix]
How easy users can play with data and analyze their
change – show the ergonomic properties of visualization in-
terface. Homogeneous occupation of visual layout, edgeless,
continuity in changes should feature good visualization [23].
Several papers described particularly this application from
the end-users-analysts point of view [10], [24]. Another
practical aspect pertain relevant documents retrieval due to
visual representation. This still remains the main direction of
current study. After receiving good precision for a small
sample, authors intend to repeat experiment with bigger
dataset and all presented methods.
Recent research [20],[21] show a strong correlation be-
tween the FD and lacunarity. To check this, it is required to
have a more representative dataset i.e. be multi-various. To
find essential changes including paradigms, the period of
analysis must be extended to three or four decades, i.e. until
2017. There basic technological problem has appeared: the
ACM has changed the classification and applied it to the col-
lection of 2013. The success to supplement the dataset de-
pends on whether the ACM will standardize the old classifi-
cation schemes according the new version and adapt it to the
whole dataset.
Undoubtedly, a sequential series of three maps is not
enough to estimate knowledge evolution dynamics. A more
multi-variety dataset to track all changes in fast growing
knowledge is needed, but truly objective circumstances con-
cerning data gathering were appeared. However, proposed
measures can be considered if we need to select the best data
distribution in the terms of interface functionality.
VI. SUMMARY
Visual interface for analyzing how computer science
evolved through the two last decades is briefly presented in
current paper. This application includes an interactive 3D
map of scientific articles organized by their semantic rela-
tionships. The authors proposed the conception how to quan-
titatively evaluate different visualization maps in respect of
possibilities of dynamics analysis. They also characterize
topological arrangement in the terms of navigation function-
ality.
Four methods of mapping including own approach of
mapping (MDS with spherical topology) have been com-
pared. Quantitative measures allowed selecting the most ap-
propriate mapping way with an accurate reflection of the cur-
rent changes of computer science. In the quantitative analysis
authors tracked the changes of pattern clusterization over
time. Clusterization degree they evaluate using machine
learning and pattern recognition algorithms (Table 1). They
adopted both lacunarity and the fractal dimension of visual-
ization patterns to find the scale of randomness in dynamics
(Table 2). Moreover the local density differences, clusteriza-
tion, rarefaction, homogeneity, and gappiness were measured
to show the most acceptable layout for analysis, perception
and exploration processes. 3D MDS maps (authors’ ap-
proach) and SOM have shown the better properties than
VOS and VxOrd. These results have proved the findings and
interpretations obtained from qualitative analysis [9]. Given
maps have revealed essential changes in computer science
literature during the time of the development of the CCS
classification compared.
TABLE II. LACUNARITY AND FD FOR EACH VISUALIZATION MAP.
Method Lacunarity Fractal Dimension
1988 1998 2009 1988 1998 2009
Authors’ method 0.0185 0.0155 0.0147 2.34 2.39 2.50
VOS 0.0035 0.0144 0.0067 2.23 2.15 2.40
VxOrd 0.0247 0.054 0.0421 2.18 2.15 2.23
SOM 0.3340 0.305 0.319 1.82 1.84 1.97
VESLAVA OSINSKA, ADAM JOZWIK, GRZEGORZ OSINSKI: MAPPING EVALUATION FOR SEMANTIC BROWSING 345

APPENDIX
346 PREPRINTS OF THE FEDCSIS. Ł ´
OD ´
Z, 2015

REFERENCES
[1] V. Osinska and P. Bala, “Classification Visualization across Mapping
on a Sphere”, in: New trends of multimedia and Network Information
Systems. Amsterdam: IOS Press, pp. 95-107, 2008. ISBN 978-1-
58603-904-2.
[2] V. Osinska, P. Bala and M. Gawarkiewicz, “Information Retrieval
across Information Visualization”. IEEE Xplore Digital Library:
Proceeedings of 2012 Federated Conference on Computer Science
and Information (FedCSIS), Wroclaw, 2012, pp. 233 – 239.
[3] K. W. Boyack, B. N. Wylie and G.S.Davidson. “Domain visualization
using VxInsight for science and technology management”. Journal of
the American Society for Information Science and Technology, 53(9):
764-774, 2002. doi: 10.1002/asi.10066.
[4] Ch. Chen, Information Visualization. Beyond the Horizon. 2nd ed.
London: Springer, 2006, pp.143-170. ISBN: 978-1-84628-579-0.
[5] K. W. Boyack, R. Klavans and K. Börner, “Mapping the backbone of
science”, Scientometrics, vol. 64(3): 351-374, 2005. doi: 10.1007/
s11192-005-0255-6.
[6] N. J. Van Eck and L. Waltman, “VOS: a new method for visualizing
similarities between objects”, in Advances in Data Analysis:
Proceedings of the 30th Annual Conference of the German
Classification Society (eds HJ Lenz, R Decker), London: Springer, pp.
299-306, 2007
[7] N. J. Van Eck and L. Waltman, “How to normalize cooccurrence data?
An analysis of some well-known similarity measures”, Journal of the
American Society for Information Science and Technology, 60(8):
1635-1651,2009. doi: 10.1002/asi.21075.
[8] N. J. Van Eck, L. Waltman, R. Dekker and J. Van den Berg, “A
comparison of two techniques for bibliometric mapping:
Multidimensional scaling and VOS”. Journal of the American Society
for Information Science and Technology, 61(12): 2405-2416, 2010.
[9] F. Moya-Anegón, V. Herrero-Solana and E. Jiménez-Contreras. “A
connectionist and multivariate approach to science maps: the SOM,
clustering and MDS applied to library and information science
research”. Journal of Information Science, 32(1): 63-77, 2006.
doi:10.1177/0165551506059226.
[10] V. Osinska and P. Bala, “Study of dynamics of structured knowledge:
Qualitative analysis of different mapping approaches”, Journal of
Information Science, 1-12, 2014. doi: 10.1177/0165551514559897.
[11] S. Fortunato, “Community detection in Graphs”, Physics Reports,
486: 75-174, 2010. doi: 10.1016/j.physrep.2009.11.002.
[12] C. Ware, Information Visualization: Perception for Design. CA:
Morgan Kaufmann, pp. 11, 188, 273, 2004. ISBN 0123814642.
[13] J. D. Banfield and A.E. Raftery AE, “Model-based gaussian and non-
gaussian clustering”, Biometrics, 49: 803-821, 1993. doi:
10.1093/biomet/63.3.413.
[14] C. Biernacki, G. Celeux and G. Govaert, “Assessing a mixture model
for clustering with the integrated completed likelihood”, IEEE
Transactions on Pattern Analysis and Machine Intelligence, 22: 719-
725, 2000.
[15] P. A. Devijver and J. Kittler, Pattern recognition. A statistical
approach, London: Prentice Hall, 1982.
[16] R. O. Duda, P. E. Hart and D. G. Stork, Pattern classification, New
York: John Wiley & Sons, 2001.
[17] A. Jozwik, “A learning scheme for a fuzzy k-NN rule”, Pattern
Recognition Letters, 1: 287-289, 1983. doi: 10.1016/0167-
8655(83)90064-8.
[18] A. Jozwik, S. Serpico and F. Roli, “A parallel network of modified 1-
NN and k-NN classifiers -application to remote-sensing image
classification”, Pattern Recognition Letters, 19: 57-62, 1998.
[19] R. E. Plotnick, R. H. Gardner and R. W. O'Neill “Lacunarity indices
as measures of landscape texture”, Landscape Ecology, 8(3): 201-
211, 1993.
[20] A. Forsythe et al., “Predicting Beaty: Fractal dimension and Visual
complexity in art”, British Journal of Psychology, 102, 49-70,2011.
T. G. Smith, G. D. Lange and W.B.Marks, “Fractal Methods and
Results in Cellular Morphology”, Journal of Neuroscience Methods,
69: 1123-126, 1996. doi: 10.1016/S0165-0270(96)00080-5.
[21] V. Osinska, “Fractal analysis of Knowledge Organization in Digital
Library”, in Katsirikou A, Skiadas CH (eds) New Trends in
Qualitative and Quantitative Methods in Libraries, Singapore: World
Scientific Publishing, pp. 17-23, 2011.
[22] W. A. Pike et al., “The Science of Interaction”, Information
Visualization, vol. 8, 4: pp. 263-274, 2009.
[23] V. Osinska, J. Dreszer-Drogorob, G. Osinski and M. Gawarkiewicz
“Cognitive Approach in Classification Visualization. End-Users
Study”, in Classification & Visualization: interfaces to knowledge (ed
A. Slavic et al), Hague, Holland, 23 -25 October 2013, Würzburg:
Ergon Verlag, pp. 273-283. ISBN 978-3-95650-007-7.
VESLAVA OSINSKA, ADAM JOZWIK, GRZEGORZ OSINSKI: MAPPING EVALUATION FOR SEMANTIC BROWSING 347