=Paper=
{{Paper
|id=Vol-2378/longBDE2
|storemode=property
|title=
An Approach to Visualize Implications
|pdfUrl=https://ceur-ws.org/Vol-2378/longBDE2.pdf
|volume=Vol-2378
|authors= Pablo Cordero, Manuel Enciso, Ángel Mora, Pablo Gomez González
|dblpUrl=https://dblp.org/rec/conf/icfca/CorderoE0G19
}}
==
An Approach to Visualize Implications==
https://ceur-ws.org/Vol-2378/longBDE2.pdf
An Approach to Visualize Implications
Pablo Cordero, Manuel Enciso, Ángel Mora, Pablo Gómez González
Universidad de Málaga (Spain)
pcordero@uma.es, enciso@lcc.uma.es, amora@ctima.uma.es
Abstract. Visualizing implications has become a hot topic, providing
new solutions to reveal the knowledge contained in the sets of rules. In Big
Data applications, it is even more fundamental since their datasets usually
produce a huge number of rules. Data visualization can be considered as a
tool to illuminate the information, guiding the search for important rules
or significant attributes. Usually, the user do not want to exhaustively
examine all the implications, but rather to analyze the relevant knowledge
in the rules.
In Formal Concept Analysis (FCA), some well-known tools allow to visu-
alize the concept lattice and, even more, the implications. They focus on
how to present these two visions of the same information, but they do not
extract a further knowledge. Here, we present a new visualization model
for implications, oriented to display and infer some interesting insights
from the set of implications.
Keywords: Data visualization · implications · formal concept analysis
1 Introduction
Big Data has quickly become a popular topic and this situation has inspired
many researchers to provide new methods, techniques and solutions to address
some of its challenges. In [6] the authors make a review of Big Data applications,
challenges, techniques and technologies. They remark that it “also arises with
many challenges, such as difficulties in data capture, data storage, data analysis
and data visualization”. In this work we focus on this last challenge, which is
closely tied, in our opinion, with the data analysis issue. In big data, where the
high volume does not only involve the amount of data but also its complexity,
data analysis has to be improved with some way to visualize the information,
providing the human interaction too.
In [25], the authors summarize the power of data visualization, outstand-
ing its benefits in different issues, adding a percentage measuring their impact.
In the top three of this study they have highlighted the following: Improved
decision-making (77%), Better ad-hoc data analysis (43%) and Improved collab-
oration/information sharing (41%).
Copyright c 2019 for this paper by its authors. Copying permitted for private and
academic purposes.
2 P. Cordero et al.
Data visualization has been traditionally conceived to be the last step of data
analysis. However, when data analysis refers to big data, data visualization allows
the user to discover more insights by himself, joining the power of big data tech-
niques and the human interaction [22]. This view of data visualization has been
strongly emphasized by some authors [6]: “The interactive analysis processes the
data in an interactive environment, allowing users to undertake their own analy-
sis of information. The user is directly connected to the computer and hence can
interact with it in real time. The data can be reviewed, compared and analyzed in
tabular or graphic format or both at the same time.”
Data visualization is not a new trend in FCA. Usually, the Concept Lattice is
represented as a Hasse diagram where some elements has been added to enrich
its semantics, adopting the so-called line diagram [10]. Since the very beginning,
it was used to represent the concept lattice in a proper and direct way, by using a
natural description that communicates the knowledge in a natural way [7]. This
lattice can be efficiently represented as a labelled directed acyclic graph where
vertices are formal concepts and its adjacency relation is the transitive reduction
of the ordering relation. The line diagram has a strong point: it enclose several
views of the information (concept order relation, implications, intent/extent con-
nection, etc) in just one –and simple– representation. However, in our opinion,
behind its apparent simplicity, the user has to be very familiarized with its in-
terpretation to properly understand all the information. In [8] the line diagram
is presented as a key element for some FCA topics: concept hierachy, attributes
partition, etc. In [21] the authors study this diagram and make emphasis on three
issues to improve this data visualization approach: reduction, layout and interac-
tion. They also have reviewed three tools to visualize concept lattices, developed
by the Defense Science and Technology Organisation [20]: Carve, SORTeD and
DAnCe. In [14] the author presents an approach enriching the traditional rep-
resentation of Concept Lattices by providing new visual elements and providing
some interactive features, updating the classical representation to the new trends
in visualization. He also studies the situation of the diagram when some changes
are performed in the information, for instance when a new attribute/object are
introduced or deleted. In addition, he also presents various strategies for pruning
a formal concept lattice, to gain a clearer structure or to emphasize on interesting
parts. These strategies provide more readable diagrams than the classical one, but
visualization models are not significantly changed. In [17] the authors introduce
a significant change, simplifying the Hasse diagram by a tree-like approach. They
suggest that this variation allows a better understanding of the diagram, since
the simplification from graphs to trees preserves all lattices entities and allows
the user to focus on some structures and patterns, better guiding the search of
insights. In addition, they enriched the visual vocabulary by adding colors and
size, both in the nodes and in the lines, to add some semantic information to the
new model. We end this review with the work presented in [1], where the authors
mainly use Hasse diagrams to visualize concept lattice but with a strong root on
the FCA framework. Hence, they provide the visualization of Pattern Structures
and AOC-posets, concept annotation, filtering concept lattice and pattern concept
An Approach to Visualize Implications 3
lattice based on several criteria. Filtering allows the representation of sublattices
and superlattices of interesting concepts, providing a way to easily discover new
insights.
Unlike the others works we have just previously mentioned in this review, this
work also includes a former visualization of implications. Traditionally, FCA liter-
ature has accepted this two fold views of the same knowledge, designing methods
to represent and navigate in the concept lattice and other ones to manage and
infer new implications. Recently, some authors have proposed the use of data vi-
sualization to represent implications and association rules. This issue is far to be
finally addressed. In this work we propose a new step to approach the implication
visualization and exploration. Our starting point is a set of implications, no mat-
ter how it has been extracted from the data set. We refer the reader interested in
this area to the recent works [3,19].
The work is organized as follows: in Section 2 we make a review of the tools that
have been presented to visualize the implications, with a particular emphasis on
the R package arulesViz presented in [11]. In Section 3 we present the contribution
of our work: a new model to visualize the implications. The model has been guided
by some semantics characteristics of the attributes, in particular by the role that
the attribute plays in the system. We end the paper with a Conclusions and
Future Works section.
2 Visualization of Implications
In our work, we focus on the implications appearing in Formal Concept Analysis
framework. As we have said, in [1] the authors propose an interesting tool to visu-
alize the concept lattice and they also add some visualization of the implications
associated to the concepts. This work can be considered as a first step in FCA.
Unlike the authors put their attention on the concept lattice, a first approach
regarding implications is presented and a basic visualization of implications with
some direct information is showed. The authors also make a brief reflection on the
need to delve deeper into this issue [2]. They include in LatViz 1 two preliminar
visualizations:
– One plot with the ID of the each implication and the attributes appearing in
their left hand side (LHS) and the right hand side (RHS).
– One matrix where in the rows appear the LHS and in the columns the RHS,
so that the matrix is a kind of relationship where each element corresponds
to an implication. Some colors are used to show the support and lift.
The authors conclude that this approach is not suitable for large sets of impli-
cations. They also provide a third diagram, a scatter plot where each dot corre-
sponds to an implication. The x-axis is the increasing support and the y-axis is
the increasing lift.
1
The plots described here can be accesible in the cited works, but unfortunately they
cannot be generated directly with the tool (https://latviz.loria.fr/).
4 P. Cordero et al.
In the literature, there exist other tools to display the information about rules:
WiFisViz [15] and VisAR [23]. In our opinion, both of them present a very similar
approach that the LatViz attribute matrix already described.
We now review the R package arulesViz [12,13], which can be considered as
a proper tool to visualize association rules. It is oriented to present the infor-
mation so that visual analysis can be further developed. As M. Hahsler says in
[12]“mining association rules often results in a vast number of found rules, leaving
the analyst with the task to go through a large set of rules to identify interesting
ones. Visualization and especially interactive visualization has a long history of
making large amounts of data better accessible”.
arulesViz incorporates the usual elements to visualize rules, already used in
the tools we have mentioned above, and it also adds some clustering techniques.
It allows to show the rules in some flexible ways as follows:
– A detailed information is visualized by means of a scatter plot called “two-key
plot” [24] where support and confidence are used for the x and y-axes and the
color of the points is used to indicate the number of attributes contained in
the rule (the order).
– As a second level in the visualization, a grouping technique is used to design
a new method, called grouped matrix-based visualization, which is based on
a method of clustering rules.
– Using a graph-based plot to connect the attributes appearing in the associa-
tion rules.
One of the interesting features of this tool is the incorporation of some cus-
tomization to perform some kind of visual analysis. Thus, it is possible to change
the information in the axes and it also allows us to show a third or fourth pa-
rameter by means of other visualization elements like color or shadow. These
elements provide a good balance between simplicity of the plots and quantity of
information [9].
We refer the reader to [13] for further details of these features. In the rest of
the section, we summarize how arulesViz works to extract interesting knowledge.
The starting point is to use the R language to generate the rules [11] using the
apriori algorithm 2 . Then, these rules are directly depicted in a scatter plot (see
Figure 1) where each rule is a dot, having two measures on the axes (support and
confidence) and a third one (lift) by using color intensity 3 . This plot provides a
great amount of information, in a very intuitive way. However, the information
displayed is centered in the rule as an atomic element, which must to be improved
for a better knowledge discovery.
In the next plot (see Figure 2), the idea is to show some semantic information:
the role played by the attributes in the rules. Thus, the different itemsets in each
2
As we previously mentioned, we use the extraction method included in arulesViz, but
any other method can be used istead.
3
In this section we are visualizing the set of rules of the Adults dataset in the UCI
Repository (http://archive.ics.uci.edu/ml/datasets/adult) where the threshold sup-
port is set to 0.3 and the confidence to 0.55. This set has 563 rules.
An Approach to Visualize Implications 5
Fig. 1. Basic arulesViz scatter plot
antecedent and consequent are depicted in the X and Y axes respectively. Both
index sets are independent, thus the same number do not represent the same
itemset. An interactivity feature is included in arulesViz: when a dot is clicked
in the plot, the full description of the rule, including the name of the attributes
in the LHS and RHS, is showed. Rules can be colored according to the value of
a third feature (in this case, we have selected their lifts). The problem with this
plot is that for a huge number of rules, it would be impossible to discover some
useful insights. Even for a mid size volume of rules, the connection between the
premise and the conclusion of the rules is not easily identified in the plot.
A third plot, named the Two-key plot [24] is also included in the arulesViz
package (see Figure 3). In the axes, support and confidence are showed and the
color of the points is used to indicate the number of attributes in the rule. This plot
not only shows the usual parameters of the rules (support, confidence, etc.) but
it also incorporates some properties that can help to identify the most important
rules when a huge number of rules is visualized.
As we mentioned in the Introduction, the current trends in data visualization
includes some kind of user interaction [22] so that it provides not only a way
to communicate the data analysis we have already did, but also to make some
further exploration by means of the visual analysis. In arulesViz, some interactive
features for scatter plots have been considered, and you can do zoom into a plot,
panning, and hovering over points to obtain a deeper information about the rule.
In addition, it also allows rule inspection by selecting a region or a point in the
6 P. Cordero et al.
Fig. 2. Representation of LHS and RHS itemset matrix in arulesViz
plot, filtering rules, etc. To end the review of arulesViz, we remark that the author
consider a limit of 1,000 rules to be displayed [13].
As a general conclusion to this section, in our opinion the visualization models
designed for implications are extremely centered on the implications itself. Thus,
they are focussed on displaying the rules and their natural characteristics: support,
confidence, lift, attributes in the RHS or LHS. However, when the number of
implications grows, this information is difficult to be visually analyzed and some
other elements need to be displayed. In our opinion, this new elements have to be
defined so that interesting insights allows to infer the role that the attributes play
in the system. This knowledge seems to be difficult to extract from the parameters
usually included in the visualizations. Thus, we present a novel approach in this
line in the following section.
3 Our Approach to Data Visualization for Implications
As we have explained, strategies presented in literature were strongly based on
the use of some basic information that could be interesting when the set of impli-
cations has a small or medium size. However, when its size grows this information
does not provide interesting insights to make visual analysis. In this situation, the
visualization has to propose new models based on some semantics information or
patterns to discover interesting knowledge.
In this work, we propose a three-step visualization model to develop a visual
analysis process. We approach this model by designing two interactive plots and
An Approach to Visualize Implications 7
Fig. 3. The two-key plot in arulesViz
a final data table. This process will help the user to move from a general set
of implications to a smaller subset. The search is guided by the role that the
attributes play in the system.
In our approach we focus on implications, but not in association rules. We
use the basic elements of association rules that also make sense for implications:
LHS and RHS. In addition, we have also designed specific parameters to enrich
the visualization model, providing a deeper knowledge of the system. The key
elements used in our visual model are the following:
– Attribute closure. We are interested in examining the full semantic power of
the premises, thus we transform the original set of implications by including
in the RHS the maximum attribute set for each premise, i.e. given a set of
implications Σ and an implication A → B ∈ Σ we built A → A+ −A where A+
is the syntactic closure of A with respect to Σ. This syntactic closure can be
approached by using Armstrong Axioms [4] or Simplification Logic [18]. This
transformation can be done in an efficient time, since the syntactic closure
has a quadratic cost on the set of implications [16].
– RHS and LHS cardinals. We count the number of attributes in the LHS and
the RHS since we are interested in the relation between the sizes of both sides
of the implication. This parameter shows whether or not the implications are
balanced. Depending of the environments, the user can search for implications
where a small itemset determines a big one or vice versa.
– Presence in RHS and Presence in LHS. We count how much an attribute
appears in all the implications by measuring the percentage of the implications
where the attribute appears, respectively, on the right and left hand sides.
– Global presence. Its is defined as the sum of the relative presence in RHS and
relative presence in LHS. Thus, we are searching for interesting attributes
instead of interesting rules.
8 P. Cordero et al.
– Generator/generated attribute roles. An attribute whose relative presence in
LHS is greater than its relative presence in LHS will be called a generator.
Otherwise, an attribute presence in the RHS is greater than its corresponding
presence in LHS is considered a generated attribute.
In the next three subsections, we describe how these parameters are used to
design the plots that are the base of our three step analysis process.
3.1 The Atribute Plot
As we have seen in previous sections, most of the works in the literature center
the visualization on the implications. However, we start our analysis with the
attributes, by plotting them according to some semantic information. Particularly,
we show the role played by each of them: is it a generator or generated attribute?
Thus, we design a plot where the axes represent the presence of the attributes
in the LHS (X axis) and the RHS (Y axis) of all the implications (see Figure 4).
In this plot, we also draw a straight line which represents the identity function.
The points located above the line are attributes playing the role of a generated
attribute. Conversely, the points bellow the line are generators. On the one hand,
a direct interpretation of this plot is the usual presence of the attribute in the
premises or the conclusions of the implications. On the other hand, this plot also
helps to figure out how the attributes are located in the concept lattice. Top and
bottom attributes in the lattice are the attributes placed close to the axes. Thus,
the attributes in the lower right quadrant of the diagram appears in closed sets
belonging to the lower levels of the concept lattice whereas the upper left dots
corresponds with attributes in the closed sets located in the upper levels of the
lattice.
Fig. 4. Atribute Plot with generated/generator attribute roles.
An Approach to Visualize Implications 9
In addition, we add some interactive capabilities to this diagram. The user
can click on any point, coloring the attribute red, and triggering a new diagram
which will be described in the following subsection. Moreover, since in some cases
the attributes come from an original many-valued formal context [5] that were
converted into a classical one by multiplication of the values. If applicable, we add
some element to visualize this situation. So, we consider that all the attributes
that corresponds with different values of an original attribute form a category.
When a dot is clicked, all the attributes belonging to the same category are
highlighted in blue color. As an illustration of this mode, in Figure 4, the red
point represents “occupation: transport moving”, while the blue ones represent
the other occupations: “occupation: Armed forces”,“occupation: tech-support”
and “occupation: craft repair”.
3.2 The Implication Plot
Once you have selected one attribute in the attribute plot, another diagram is
created. Thus, this new plot presents the information filtered according to the
selection made in the first step. In this second diagram, we shift from the at-
tribute to the implication view. Our goal now is to visualize the information of
the implications following the same idea behind the attribute plot: the balance
on the number of attributes between the RHS and the LHS. Thus, this new plot
shows the implications gathered by the cardinal of their LHS and RHS. Thus, our
idea is to group in the same dot all the implications following the same structural
schema. It also adds some additional information like the number of implications
in the group and the combination of cardinals in both sides. An example of this
diagram is showed in Figure 5. It follows the following visualization model:
Fig. 5. Implication diagram with information of the selected attribute.
10 P. Cordero et al.
– The X axis is the increasing LHS cardinal while the Y axis is the increasing
RHS cardinal. As we mentioned, each plot group all the implications with the
corresponding cardinals.
– The size of each dot represents the number of implications with the same pair
of values in both sides. The size is proportional to the number of rules for the
selected point.
– Each dot embebed an inner point representing the percentage of the implica-
tions having a given LHS and RHS size where the selected attribute appears.
– Finally, the dot color represents the main role that the attribute plays in
the implications with a given LHS and RHS. The generator character is
highlighted in red color whereas the blue color is used for the generated one.
The intermediate situations are colored in a color determined by the range
between the red and blue colors.
This diagram allows to identify if the attribute plays the generated/generator
role in an implication schema, according with their size, LHS and RHS cardinals.
Once we have discovered some insights, then we can finally get the information of
all the implications that have been grouped in each dot by clicking on the corre-
sponding dot. This will be showed in a table explained in the following subsection.
In the future, we plan to add to this table some kind of visual representation, fol-
lowing the models presented in previous works, as a help to ease the exploration
when the size of the output implication set increases.
3.3 The Rules Data Table
The final step is to show all the implications with the selected LHS and RHS
cardinals. They are placed in a data table with some classical estimators so that
the user can examine the set according to the search he has performed by using
the visual tool. This data table have been built by the user in just two actions,
the selection of one attribute and one implication schema, allowing the user to
develop visual analytics in a few steps. Following the example of the previous
section, if the points with the biggest RHS are selected –the (2,3) and (3,3) pairs–
the implications collected in these points are showed in Figure 6. This table also
has some interactive skills, allowing the user to sort it by the desired parameter.
4 Conclusion and Future Works
In the work we focus on the need to visualize the implications, which are key
elements in FCA and have not attracted as much attention as concept lattice
in this issue. We present a new visualization model designed to better discover
insights from the set of implications that was not included in other previous
methods. Our model is guided by the idea of identifying the role that the attributes
play in the implications. We have described some elements where the model is
based on and the diagrams and tables used to visually present the information.
As future works, we plan to connect this model with the visualization of the
concept lattice so that the two-fold representation that was originally proposed
in the FCA framework can be also handle in the visualization issue.
An Approach to Visualize Implications 11
Fig. 6. Information of the implications associated with the selected attribute.
Acknowledgment
This work has been partially supported by the project TIN2017-89023-P of the
Science and Innovation Ministry of Spain, co-funded by the European Regional
Development Fund (ERDF).
References
1. Alam, M., Le, T.N.N., Napoli, A.: Latviz: A new practical tool for performing in-
teractive exploration over concept lattices. In: Huchard, M., Kuznetsov, S. (eds.)
Proceedings of the Thirteenth International Conference on Concept Lattices and
Their Applications, Moscow, Russia, July 18-22, 2016. CEUR Workshop Proceed-
ings, vol. 1624, pp. 9–20. CEUR-WS.org (2016), http://ceur-ws.org/Vol-1624/
paper1.pdf
2. Alam, M., Le, T.N.N., Napoli, A.: Steps towards interactive formal concept anal-
ysis with latviz. In: Kuznetsov, S.O., Napoli, A., Rudolph, S. (eds.) Proceedings
of the 5th International Workshop ”What can FCA do for Artificial Intelligence”?
co-located with the European Conference on Artificial Intelligence, FCA4AI@ECAI
2016, The Hague, the Netherlands, August 30, 2016. CEUR Workshop Proceed-
ings, vol. 1703, pp. 51–62. CEUR-WS.org (2016), http://ceur-ws.org/Vol-1703/
paper7.pdf
3. Andrews, S.: In-close, a fast algorithm for computing formal concepts. In: Interna-
tional Conference on Conceptual Structures (ICCS) (2009)
4. Armstrong, W.W.: Dependency structures of data base relationships. In: IFIP
Congress. pp. 580–583 (1974)
5. Belohlavek, R., Vychodil, V.: What is a fuzzy concept lattice? In: Belohlavek, R.,
Snasel, V. (eds.) Proceedings of the CLA 2005 International Workshop on Concept
Lattices and their Applications, Olomouc, Czech Republic, September 7-9, 2005.
12 P. Cordero et al.
CEUR Workshop Proceedings, vol. 162, pp. 34–45. CEUR-WS.org (2005), http:
//ceur-ws.org/Vol-162/
6. Chen, C.P., Zhang, C.Y.: Data-intensive applications, challenges, techniques and
technologies: A survey on big data. Information Sciences 275, 314 – 347 (2014).
https://doi.org/https://doi.org/10.1016/j.ins.2014.01.015
7. Eklund, P., Ducrou, J., Brawn, P.: Concept lattices for information visualization:
Can novices read line-diagrams? In: Eklund, P. (ed.) International Formal Concept
Analysis. pp. 57–73 (2004)
8. Eklund, P., Villerd, J.: A survey of hybrid representations of concept lattices in
conceptual knowledge processing. In: Kwuida, L., Sertkaya, B. (eds.) International
Formal Concept Analysis. pp. 296–311 (2010)
9. Few, S.: Show Me the Numbers: Designing Tables and Graphs to Enlighten. Ana-
lytics Press, second edn. (6 2012)
10. Ganter, B., Obiedkov, S.: Attribute exploration. in: Conceptual exploration (2016)
11. Hahsler, M., Grun, B., Hornik, K.: arules - a computational environment for mining
association rules and frequent item sets. J Stat Softw 14, 1–25 (2005)
12. Hahsler, M.: arulesviz: Interactive visualization of association rules with r. The R
Journal 9/2, 163–175 (2017)
13. Hahsler, M., Karpienko, R.: Visualizing association rules in hierarchical groups. J
Bus Econ 87, 317–335 (2017)
14. Kriegel, F.: Visualization of Conceptual Data with Methods of Formal Concept
Analysis. Ph.D. thesis, Technische Universitat Dresden (10 2013)
15. Leung, K.S., Leung, C.K.S., P. Irani, P., L. Carmichael, C.: Wifisviz: Effective visu-
alization of frequent itemsets. 2008 Eighth IEEE International Conference on Data
Mining pp. 875 – 880 (2008)
16. Maier, D.: Theory of Relational Databases. Computer Science Pr (1983)
17. Melo, C., Grand, B.L., Aufaure, M., Bezerianos, A.: Extracting and visu-
alising tree-like structures from concept lattices. Proceedings of the Inter-
national Conference on Information Visualisation pp. 261 – 266 (08 2011).
https://doi.org/10.1109/IV.2011.46
18. Mora, A., Cordero, P., Enciso, M., Fortes, I., Aguilera, G.: Closure via functional
dependence simplification. Int. J. Comput. Math. 89(4), 510–526 (2012)
19. Outrata, J., Vychodil, V.: Fast algorithm for computing fixpoints of galois connec-
tions induced by object-attribute relational data. Information Sciences 185(1), 114
– 127 (2012)
20. Pattison, T.: Interactive visualisation of formal concept lattices. CEUR Workshop
Proceedings 1244, 78–89 (01 2014)
21. Pattison, T., Weber, D., Ceglar, A.: Enhancing layout and interaction in formal con-
cept analysis. In: 2014 IEEE Pacific Visualization Symposium. pp. 248–252 (March
2014). https://doi.org/10.1109/PacificVis.2014.21
22. Simon, P.: The Visual Organization: Data Visualization, Big Data, and the Quest
for Better Decisions. Harvard Business Review (6 2014)
23. Techapichetvanich, K., Datta, A.: Visar : A new technique for visualizing mined
association rules. LNAI 3584, 88 – 95 (2008)
24. Unwin, A., Hofmann, H., Bernt., K.: The twokey plot for multiple association rules
control. PKDD01: Proceedings of the 5th European Conference on Principles of
Data Mining and Knowledge Discovery pp. 472–483 (2001)
25. Venkataramana, S., Subash, S.R., Prakash, P.: Visualization of big data: Its tools
and challenges. International Journal of Applied Engineering Research 9, 5277–5290
(01 2014)