Many
existing approaches to studying and analyzing data of various origins and
complexities have a critical dependence on the quality of this data
(completeness, reliability, errors) [1], [2]. In the absence of complete data
volume, for example, the dynamic system being investigated by the user, these
issues can create difficulties and necessitate adjustments to the research
methodology. The adjustments may be aimed at clarifying research task or at
compensating for existing challenges due to the use of resources that were not
previously employed in the analysis process.
Examples of
expanding data analytics capabilities through an instrumental approach include
a variety of visual research techniques. The purpose of these tools is the
increase of analytical process effectiveness through the efficient combination
of computational, information, and cognitive resources available to researchers
[3], [4], [5]. The means of research in these techniques are visual data models
that may differ both in variants of the employed visualization metaphors and
methods of communication between users and initial data. This is implemented
using interface elements of visual model. Thereby, the visual data model
function as an interactive high-tech tool for solving data analysis problems,
both as an autonomous tool and as a component of information system [6], [7],
[8].
The partial lack
of values in the initial data may be the result of different reasons. In
general, any values may be attributed to the missing data if they raise doubts
or contradict other parameters. In this case, research task is divided into two
stages: a preliminary examination of initial data and subsequent analysis. At
the second stage, issues of assessing the significance of missing data and
reconstructing the necessary information elements are raised. An attitude is
also formed towards the results of further analysis techniques applied to
studying the initial data volume and data supplemented with the results of
recovery process. The objects of this work are visualization tools that allow
to find answers to questions from both stages, or, at least, to conduct an
expert assessment of the initial data.
During the visual
model formation for incomplete data, it is possible to rely on the following
heuristics: the value of one object within the analyzed sample of parameter
falls within the value range and characterize this parameter, but it belongs to
other objects. In other words, the uncertainty of missing values is limited (to
some precision) by values present in the initial data.
This assumption
imposes significant limitations on the range of issues that can determine the
visual data analysis task. The purpose of analysis cannot be the search for
abnormal values, extreme points, errors, etc. The most justified direction for
efforts of visualization tool developers should be considered the design of
tools that allow to perform analytics and form a holistic view of studied data.
This representation is understood as a unified object of perception – a visual
model. Its interpretation guides the analyst to the reasons for appearance of
specific values in the data under study. For example, a visual model of
multidimensional data, based on the idea of parallel coordinates, for
representing initial data in a single space describes the current state of
studied system, with parameters containing descriptions of over 300 informative
elements. The representation metaphor makes it easy to identify objects with
similar properties, while the degree of "proximity" and its criteria
are determined only visually (Fig.1).
Fig.
1 A visual model of multidimensional data using the idea of parallel
coordinates
Some assumptions
about properties of such an image, which acts as a model object and is intended
for studying and conducting cognitive research, are acceptable [9], [10]. Sets
of experimental data characterizing objects of studied field may differ in
terms of their volume, acquisition conditions, and states of the object systems
or their individual properties, reliability, etc. During preliminary research,
the analyst's task may be in obtaining general value judgments that do not
contradict either the available data or those that may appear later. Any
redundant information in this situation complicates researcher's work. Therefore,
the requirement for a visual model is to simplify perception of the overall
data picture without losing its informational content.
Within the
framework of the analytical solvable task, determining elements of source data
or their representation that have excessive information content is an
independent issue. The answer to this question becomes an essential part of
interpreting initial data. In the developed methodology of visual research,
this stage is mandatory and its passage is ensured by the availability of
interactive features for visual analysis tools that are necessary for user to
make a decision about excluding redundant data from the field of view. In case
the negative interpretation results are achieved, the data will be returned to
work [11], [12], [13].
In case of forming
only a general idea of the studied system, i.e. without examining secondary
data, elements of initial data may be excluded from perception: event
identifiers, variable values, or repeating scales (Fig.2). Identifiers are of
interest at the stage of comparing data sources and the features of their
changes. However, for the initial analysis this information may be relevant if
the events on which attention needs to be focused are known. An independent
simplification technique is replacing smooth segments connecting data points in
the description of each informative element with linear visual connections. In
this case, the visibility of presentation is enhanced, but analysis of the
entire volume of initial data is hindered by a large number of visible and
distinguishable informative images (data points, angles in the images of
elements that are concentrators of attention).
The values of
individual parameters may also be redundant information, since in most cases
the basis for obtaining an evaluative judgment is the dynamics of changes (or
lack thereof) of these. It can be assessed by comparing descriptions of events
[14], [15]. Consequently, the joint analysis of several event descriptions can
be more effective than the sequential research due to better concentration of
the user's attention. An experimental evaluation of the user's decision-making
speed on the significance of individual information objects when using the
parallel coordinates metaphor confirms the possibility of applying visual
models with a simplified visualization metaphor (Fig. 2). The time required to
complete the expert assessment stage is reduced by 5-15%.
Fig.
2 Variant to simplify the visualization metaphor
Visual
representation of multidimensional data inevitably generates a variety of
cognitive distortions. They are associated with the simultaneous observation of
data images in a shared visual space. Their scaling does not correspond to the
values of their own scales or units of measurement, but rather to the
peculiarities of the representation area [16], [17]. A promising solution could
be the transition to free user scaling of visual images of individual
dimensions. This would allow to exclude (temporarily or permanently) some or
all measurements in multidimensional source data from the visual field of the
scale, as well as change their scales according to the researcher’s current
needs.
Optimization of
the researcher's field of view is associated with the exclusion of secondary
data and partially limits the variety of issues available for discussion. In
particular, the analysis of values present in the source data becomes
independent of the absolute values of the parameters and is based on relative
values. A special case of this statement of the purpose for the data
examination stage may be the requirement for the researcher to form a holistic
view of the properties of the data being analyzed [18]. For example, an
evaluation of data quality as a ratio in the initial data volume of complete
and incorrect object descriptions (Fig. 3). In this example, a version of the
metaphor for visualizing multidimensional data is developed for a preliminary
assessment of the quality of investigated data. Each line corresponds to the
image of a multidimensional information object, represented in the
"spherical" version of parallel coordinates. Data without gaps forms
a line that is completely located on the surface of the visual model
"sphere" (colored lines, each representing one information object).
Information objects with partially missing values have fewer projection points
on the spherical surface of the model, and therefore easily stand out in the
general data image (blue lines). This situation can be effectively used if
there is the possibility to automatically scale the visual representation.
However, it destroys the understanding of the range of observed value changes.
The importance of maintaining this understanding is not a requirement for
visualization tools, since the magnitude of the range of changes is typically
known or can be communicated to the researcher using minimal visual methods.
Comparative
analysis based on matching of visually observed features in the source data
reduces the total analysis time (up to 5%, depending on the data itself). A
useful result of visualization in this case is the prediction of parameter
values that are missing in the initial data. The unavailability of individual
values or significant data fragments may be the result of losses, measurement
errors, or researcher's area of interest being beyond the capabilities of
available data sources. In each of these situations, predicting missing values
is based on the peculiarities of the researcher's visual perception. Implementation
is carried out by means of visual analytics, and, importantly, based on the
user's subjective knowledge associated with experience in solving similar
analysis tasks.
The data
reconstruction performed in this way is similar to the formulation of hypotheses
explaining the origin of the entire data volume. This can be either an
independent task or an intermediate stage of a broader study. It is important
that the persuasiveness of visualization, which is one of its most significant
advantages, does not become an obstacle to critical verification of the
formulated hypotheses. A convenient, although not always possible, solution to
this issue may be a comparative analysis of not the initial data itself, but
the consequences of adopting different hypotheses visualized as alternative
options in the data model [11].
Fig.
3 A metaphor of multidimensional data visualization for preliminary assessment
of the data quality
The basis for
evaluation and necessary reduction the conflict of visualization techniques
used in the general perception space may be the allocation of independent basic
visual attributes with minimal mutual influence. In other words, it is
necessary to understand the compatibility of informative flows in the visual
perception process, which can be painlessly separated at the stage of their
interpretation. The study of applied visualization systems, including tools
used in business analytics and scientific visualization, allows to notice that
there are very few known and used ways to display mutually compatible data
[19]. It is quite easy to identify two large groups of visualization techniques
that can be used independently and jointly: interpretation of the dynamics of
changes and deviations from conditional equilibrium.
Techniques based
on the observation of dynamic of changes in the data under study (Fig. 4) are
characterized by a high concentration of observer attention and the rate of
hypothesis generation. In experimental versions of the visual model, a dynamic
metaphor was used to preliminary assess initial data characteristics. In this,
representing the image of each informative object is a sequential animated
process of the image appearance. Each point corresponds to the state of one
informational object present in initial data. Point position within the visual
model space is defined as the resulting vector in the spherical analogue of
parallel coordinates (on the right). The alternative variant (on the right) is
distinguished by a dynamic demonstration of the sequence of object states,
which significantly exceeds the static version in information content and rate
of preliminary analysis.
|
|
|
Fig.
4 Experimental versions of the visual model for a preliminary assessment of the
initial data characteristics
The general image
of initial data, which becomes more complicated over time, is well interpreted
if there is a small number of objects (less than 50), for data with a low
dimension (the number of parameters in one description is up to 10). In
general, interpreting the volume of initial data requires special user
training.
Besides, the
observation of a changing data image, in which the dynamic component may be not
only time, but also any other variable initiates and supports the user's mental
activity associated with the reconstruction of missing or otherwise unavailable
data. Experimental evaluation of this approach’s advantages is complicated by
the need to define and systematize ideas about a potential user. In some
measurements, when observing dynamic models, preliminary study time was reduced
by 9-12% compared to static visual models.
A feature of the
reconstruction process for this type of data is the equal employment by the
user to generate a new hypothesis of both the results of the current
interpretation and data previously available to the researcher. This feature
specifies the field of application for visual analytics tools and imposes
requirements on the user’s knowledge and experience. This means that the amount
of information a person can freely manipulate and work with a dynamic model has
natural limitations. Developers of visual representation metaphors should take
this into account.
The idea of
interpreting deviations has significant number of implementation versions. One
of the most popular is the assessment of spatial deviations in data images from
a given initial state. Other versions of the same idea are deviations in the
color representation from a predefined neutral state, as well as any other
difference between the instance of visualized data and what is expected. This
variety of versions determines the popularity of these visualization tools.
However, this very often leads to obtaining visualization metaphors that do not
take into account the patterns of visual perception. Examples of such erroneous
metaphors from the group with deviations may be common versions of data color
coding, when the peculiarities of color and brightness perception are not
considered, or the visual disproportion of scales for all or some parameters in
the multidimensional data array.
This determines the presence and
appearance of visual accents.
The advantage of
the deviation group representatives can be considered in obtaining a more
accurate interpretation of data presentation in the following cases: a
significant number of analyzed events persists for large data volumes; in the
presence of chaos in the values of studied parameters; difficult understanding
without the use of additional visualization techniques. This can be explained
by the lack of restrictions in the duration of observing the visual model
properties, as well as the possibility of simultaneously displaying a variety
of data in the visual field. The interpretation procedure for such an image has
significant differences from the models of dynamic metaphor group, as it
becomes possible to adapt the speed of cognitive processes to individual actual
user needs (Fig. 5). Version of the deviation metaphor is proposed to assess
the advantages of this type of visual models. This version is designed to form
a general idea of the initial data, as well as to decide on the possibility of
borrowing missing values from information objects with similar properties
(yellow and green lines). In the proposed model, differences in information
object’s properties are visualized as a deviation accumulated with distance
from the model center. Therefore, the detection of similar objects does not
cause difficulties.
Potential
practical applicability of visual models, using the metaphor of deviations, is
higher than that of data-oriented dynamic or static models. This is caused by
the effective use of natural visual matching mechanisms that are familiar to
most users and have a high speed. According to experimental estimates, reducing
the time spent by the observer on allocating objects of increased interest
based on differences in their properties from the total initial data volume or
explicitly specified "reference" values reaches 25-40%. It also
depends on the ability to match properties of a particular representation
metaphor and scale of studied deviations.
Highlighting these
two groups of data visualization techniques does not make it mandatory to
choose one of them when developing visual analytics tools. There are many
applied visualization systems that combine both approaches. For example,
dashboards in BI-systems combine different display options. They are focused on
emerging changes (KPI monitors). Techniques for representing dynamic changes in
the form of static images have also become popular (for example, the widespread
metaphor of "Japanese candlesticks" - Candlestick Chart [20]).
Presentation of information in the form of dynamically changing KPIs is
oriented on analyzing not the values themselves, but the moments and magnitudes
of their changes. Since simultaneous monitoring the value of several variables
creates a significant burden for the user [21], digital values are often
supplemented with graphical elements to reduce it. They indicate the direction
of recent changes, but it also complicates the overall image of data. The
metaphor of "Japanese candlesticks" is conceived as a compact and
rich form of data representation that is constantly changing. Therefore, it has
become a popular tool for solving forecasting tasks. The disadvantage of this
approach should be considered additional requirements for training of
specialists using these tools, as well as the limited amount of information
transmitted to the observer at one time. This is a consequence of transition to
an unusual sign system.
Fig.
5 Variant of a visual model based on deviation assessment
It worth to note
that for representatives of each specified group of visualization techniques,
as well as for hybrid variants, the statement regarding target use of visual
perception patterns is true [22]. Understanding the role of specific perception
features in searching for answers to various questions will allow development
of visual analytics tools based on understanding and planning principles of
their operation. Unfortunately, there are many examples of incorrect attitudes
towards the application of visualization tools in practice. This raises
questions about the value of visualization in general.
For the tasks of
reconstructing missing data, the applicability of visual perception patterns
has its own characteristics, and ignoring them reduces the effectiveness of
visual analytics tools. For example, the principle of perception integrity
manifests itself in the form of an observer’s unconscious desire to mentally
combine individual visual elements into a group. This occurs when there are
assumptions about the presence of a general rule defining their appearance. In
the reconstruction task, this creates conditions for formation hypotheses about
similarity between the values of individual parameters. It happens if general
descriptions of information objects in the visual model field are similar to
each other. In the absence of other grounds for recovering missing data, this
principle may provide basis for further analysis of the system under study.
The integrity of
perception can also manifest itself through assumptions about the existence of
information structures. These structures combine similar elements in the visual
field and have their own properties that also need to be studied. However, this
is not always fair or necessary (Fig. 6). For example, a visualization metaphor
is useful, if it allows to simultaneously represent both the studied data and
areas of acceptable values. This metaphor is formed as a result of interactive
model management and can lead to the perception and interpretation of a visual
structure that is not intended for this purpose. A convenient solution to this
challenge is decomposition of the visual model. This will allow studying of its
individual components. The introduction of such control changes is aimed at
obtaining additional information by the user. It is necessary to focus on
specific values (features of the initial data, the detection of which, using
traditional data pre-processing techniques, is possible if a research hypothesis
is provided). The proposed metaphor is a tool for forming statistical
hypotheses tested at the next stage of data analysis.
Fig.
6 A visual model of multidimensional data that allows simultaneous presentation
of the volume of data under study (animated tracks - images of information
objects in the space of the visual model) and variations in visualization when
making any permissible changes to the data image (blue translucent areas)
An equally
significant feature of visual perception is connected with the principle of
generalization. It largely determines the processes of recognizing objects in
visualized information or its individual fragments. In accordance with this
principle, recognition of known to the observer object occurs even in the
presence of distortions in transmitted data or other variants of information
noise. In other words, recognition of an object ends with determination of its
belonging to a specific class of objects that belongs to the user's knowledge
system. The result of generalized perception is the ability to operate with
information about not individual objects, but their corresponding classes. This
allows restoring gaps in the initial data when features of objects are combined
at the classification level. In the metaphor using the idea of parallel
coordinates (Fig. 1), a reduction in research time by 5-10% is explained
precisely by perception generalization.
The principle of
objectivity manifests itself in creating prerequisites for dividing visual
information flow into following data: relevant to the observer's current
interest; elements that belong to the environmental group according to the
criteria established by the user. In visual analysis tasks this principle
corresponds to decomposing the visual field into objects. Their parameters and
values are perceived as independent information structures. Existence of such
structures can be explained in the course of ongoing research. In addition to
the objects, there are signs of relationships between the selected structures
and informative environment that influences the state of objects of interest.
Objectivity of visual perception is the basis for persuasiveness of observed
images and, to some extent, the data used to create the image. Therefore, the
process of data analysis is initiated, since objectivity is a confirmation of
the cognizability and reproducibility of studied events.
In addition, the
opposite side of the objectivity principle is the possibility to detect objects
– informative elements of the visual model that have unusual or atypical
properties (Fig. 3). This complements results of interpreting the observed
visual model state by raising new questions and corresponds to managing the
process of analyzing initial data (including its completion or changing the
goal). In discussing data recovery tasks in descriptions of individual information
system elements, the objectivity of perception can play a significant role.
Since this is a factor in convincing the user about the reasonableness of
initial data and possibility of their extrapolation to descriptions of elements
that lack necessary values. The absence of such conviction causes repeated
cycles of verifying intermediate interpretation results. Their number exceeds
what is necessary and depends on the subjective readiness of the researcher to
accept controversial and non-obvious hypotheses.
The constancy of
visual perception becomes a pattern that is of great practical importance for
visualization tasks in general, as well as issues of using visual analytics
tools for data recovery and, in particular, data analysis. Visual representation
of multidimensional data, based on the metaphor of complex geometric images in
three-dimensional space, explicitly uses the perceptual constancy to provide
the user with ability to interactively control the model space. It also allows
the user to control the sequential change of observation points for data image
or productive refinement of metaphor. The principle of constancy ensures the
user maintains understanding of changing data model image. It is frequently
used to organize the possibility of an arbitrary direction for analyzing
multidimensional descriptions of objects. Moreover, it takes into account the
subjective distribution of interest areas in the data space.
In the process of
designing visualization tools, the principle of perception constancy imposes
restrictions on the range of allowable changes in the data image.
Transformation of the perceived data image, any systematic conversions of the
representation metaphor are independent techniques of cognitive research. They
are aimed at providing users with ability to search for interpretation results
that correspond to questions posed. However, according to the principle of
constancy, exceeding the limits of permitted transformations results in
changing the class to which an object of perception (or its individual parts)
was assigned. This may be a desirable outcome if changes in the visual model
were made to form a qualitatively new interpretation hypothesis. It can also be
a negative phenomenon if the change in class of informative object has caused
destruction of the system of facts, connections between them and the
corresponding knowledge previously acquired by user.
Considering and
applying the principles of visual perception mentioned above is an obligatory
aspect of designing visual analytics systems. This allows researchers to make
actions and conclusions systematized. It does not mean that the cognitive
process will lack such a useful component as insight. However, its achievement
becomes dependent on the accumulation by the users of their own analytical
activities results. In addition, an important stage in the cognitive process is
its completion. It refers to the state when the answer to the initial research
question has been obtained, and there is sufficient confidence in the
correctness of this answer to move on to a new question and a new research goal
of the degree. The time it takes for a user to reach this state determines the
effectiveness of tools and corresponding analysis techniques. This is
determined not only by the properties of problem being solved and visual
analytics tools, but also by the interaction of these tools with the user.
Discussion of the
specific aspects of interaction between the visualized information and
perception and researcher’s mindset cannot be limited to discussing the
applicability of classical perception patterns in the interpretation of visual
data models. This is associated with a significant increase in the variety of
expressive techniques that can be used in visualization tasks, as well as the
emergence of specific visualization tools that are related exclusively to
digital technologies and do not have close analogues in reality [23].
An example of this
trend in visualization is the use of animation. Animation is the movement of
elements, that constitute a data image composition, or any other metaphor
transformations, or changes in the work of methods included in a general
metaphor, that are synchronized in the user's perception with the time of
representation observation (Fig. 4). Evaluation the prospects of involving
animation in visual analytics tasks indicates that judgments, related to the
action of perception patterns, are largely applicable to it. However, there are
several specific manifestations of the animation’s influence on the
interpretation of visualized information, which can either enhance or seriously
reduce the activity of cognitive process. The problem of animation is associated
with active involvement of the subjective movement understanding. At the same
time, there are contradictory properties that become enhanced if visual
analytics tool is developed for collective use.
One of the
significant patterns, that are used by creators of various type animated
images, is readiness of perception for the next movement or expectation, formed
as a result of understanding events, that have already occurred. In
cinematography, this fact is used to reduce the time required for moving on to
the next scene, but in visual analytics, readiness of perception is an
indicator corresponding to the viewer's understanding of causal relationships
affecting origin of studied data. This understanding may be linked to the
user's previous knowledge or new hypothesis, which was verified by the visual
model.
Direct consequence
of the active role of informed waiting in interpreting a visual data model is
the rapid detection of inconsistencies between the expected event development
in visual field and the actual behavior of the studied system (Fig. 4).
Disturbed waiting is an effective way to detect errors in two directions: in
the initial data (it is possible to evaluate the quality of studied data) or in
the user knowledge. Additionally, the waiting formed for user in the
interpreted movement is an indicator of achieving a certain level of
understanding (confirmation of the hypothesis about data origin). Efficiency,
in this case, is explained by the fact that there is no need for additional
efforts on the part of visualization tool developers to draw attention to these
errors, since switching attention to error analysis takes place naturally for
the user.
Additional
advantage of animation is its active role in maintaining a high level of
observer involvement in the study. This refers to the influence of user's
psycho-emotional state on the process of solving a visual analysis problem. In
other words, prolonged concentration of user's attention at work leads to
slowing down or stopping the cognitive process. Animation, as part of the
visualization metaphor, can regulate user involvement through such techniques
as rhythm of perception, switching between observation and interpretation
modes, and drawing personal interest [24].
In general,
methodological approach to the problem of incomplete data research consists of
choosing one of possible solution algorithms, preparing conditions for its
implementation, and developing tools. They should fit the task and determine
means of visual analytics and qualitative evaluation of the result. The
following alternative options can be considered as two possible algorithms for
conducting research on incomplete data:
A.
Elimination
of data from the research process that are not available for study in
traditional or other ways, due to the lack of values in the part of samples.
Creating a data model based on the assumption about insignificant role of
missing data in the process of answering questions facing the user. In this case,
it is important to use visualization to form the researcher's confidence in the
ability to continue solving the problem of analyzing incomplete data sets. The
lack of confidence in the correctness of actions can be compensated at the
stage of evaluating the obtained analysis results. However, the transition to
verifying results is preceded by the hypothesis generation stage. During this
stage, the presence of user’s subjective uncertainty leads to excessive
multiple proposing of new hypotheses. This significantly slows down the
solution of analysis task. It should be noted that the potential applicability
of this solution algorithm is largely based on the objectivity and generality
of visual perception.
B.
Replacing
missing values in descriptions of informative data elements with assumed ones,
i.e. hypothetically possible values. A specific case of this solution is the
introduction of a special point - "no data available" on appropriate
scales. Next, analysis is carried out, taking into account the ability of state
of data sources when they do not generate necessary values. The basis for
reconstruction in this variant of the solution algorithm is the hypothesis that
individual values of multidimensional data are close to each other. At the same
time, the condition of sufficient value similarity in most other available
measurements must be fulfilled. This hypothesis is not obvious and can be
justified by solving a specific task by visually evaluating the similarity
between complete sets of values for available informative elements (if the
requirement of sufficient quantity in the original data set is met). In this
case, visualization also becomes a convenient instrument for justifying the
applicability of chosen reconstruction approach with quick preliminary analysis.
This happens thanks to the involvement of visual perception patterns such as
integrity and constancy.
Development of
visualization tools, based on these assumptions, is primarily aimed at defining
the effective metaphor of visual representation. Characteristic property in
this case is the ability to simultaneously display the full volume of available
data in the visual field. This is essential for forming a holistic view of the
research subject and its features for observer. At the same time, developed
metaphors act in similar way in both algorithms that are used for solving
analysis problem, i.e., they assess the remoteness (visible difference) or
proximity (similarity between images) of informative elements in a
multidimensional data space, but have opposite goals.
A.
Exclusion
algorithm requires a representation metaphor that emphasizes the assessment of
diversity. It means that in the case where the diversity of incomplete data has
negligible influence on the meaning, it is possible to make a decision about
the admissibility of their exclusion from the total volume of initial data.
B.
For the
replacement algorithm, a visualization metaphor should facilitate an effective
search for objects with similar properties in order to provide a basis for
borrowing missing data.
It should be noted
that in some cases the same visualization metaphor may be applicable for both
options, because the requirements for analysis tools are similar. In this case,
the correctness and effectiveness of their application will depend on the
user's preparedness, their understanding of the visual analysis purpose and
subjective effectiveness of the expressive methods used.
At the stage of
qualitative assessment of the choice correctness and success of using
visualization tools for solving any data analysis task, it becomes necessary to
determine the success of overcoming many objective problems and mistakes made
by users or developers. Common problems for a wide variety of data and tasks
are always the following:
A.
Representation
of non-numeric data. The traditional method of using categorical measurement
scales may cause certain difficulties in situations where several different
scales of this type are required to visualize the information element of
initial data. Technically, there are no difficulties and the formal procedure
for creating an image does not face obstacles. However, for users who want to
visually determine the degree of similarity or difference between data
elements, this question can become quite difficult, because understanding
distance in a semantic space defined by categorical scales rarely corresponds
to the user's subjective experience. In this case, the integrity of informative
element image formed in observer's perception can become the basis for
attributing it to a certain class of objects, and using categorical scales can
be replaced by assigning perceived attributes to objects.
B.
Assessment
of the significance. The ratio in the visual representation of the significance
of individual properties is both a reserve of any representation metaphor and
its weak side. In this case, it is difficult to determine in advance the
factors that control user's attention. In practice, this corresponds to a high
probability that significant elements of the data image will escape observer's
attention. Alternatively, minor fluctuations in values can cause a deep
cognitive process and distort the general meaning of studied data.
C.
Abstractness
or the metaphor complexity. Geometric properties of a visual image represent a
model picture of the source information, and are very rarely directly related
to the real properties of the data. The representation metaphor for any
multidimensional data, in the perception of an ordinary observer, does not
correlate in any way with the perception of data itself. A simple illustrative
example of this state can be a regular geographical map, where information
about the height of surface point is presented in the form of color coding. In
more complex situations, when the studied data has more complex organization or
properties what are unknown to the user, any metaphor generates an image with
high visual complexity and degree of abstraction. In general, this fact is not
problematic, but taking into account perception patterns, unexpected results
may occur when the observer determines the category of an object or its
informative properties.
Direct consequence
of these difficulties is the increasing significance of the experience,
perception and cognitive skills of a particular researcher. There is an
obligatory stage of determining the features of information communication
performed using visual models. The selection of data representation method can
be made manageable and controllable in order to delegate users the opportunity
for interactive subjective adjustment of the model to their needs or
perception.
A
promising area of application for the approach discussed in this work is
solving data processing and interpretation problems, that arise from
cyberphysical systems (CPS) of various complexity levels, operating in an
autonomous or partially controlled mode. These systems have a large number of
sensors responsible for collecting many types of data, differing both in the capacity
of corresponding information channels as well as their speed and reliability.
Examples of such cyber-physical systems/CPSs are unmanned aerial vehicles
(UAVs), robotic stations operating under conditions that prevent the
acquisition of objective observation experience (deep-sea robots), and many
other multimodal monitoring systems.
All these tasks
are characterized by the need to obtain analysis results under the following
conditions: the impossibility of taking repeated measurements, the presence of
high levels of noise, the loss of original data fragments. During evaluating
advantages of the approach considered in this work to solving the visual
research problems, a set of visual metaphors was developed. It is designed for
assessing the quality and preliminary analysis of heterogeneous data. The
source of this data is monitoring systems. The test data had different volume
and level of complexity: from 20 to 300 records, each of them could contain
5-18 parameters of different types, including measurements of temperature,
pressure, gas content, etc. (Fig. 7). Using unified metaphors in relation to
the analysis of data obtained by the CPS during environmental monitoring
allowed to reduce the time for developing visual research tools. For example,
when initial data volume increases (up to 500 records), the time required for
creating a visual model does not change. In addition, such unification has
reduced the cost of training a visual analyst to work on tasks in new
directions.
Fig.
7 An example of test data and a visual metaphor chosen for analysis
Usage of the
parallel coordinate metaphor for the task in question is caused by the need to
draw on the experience obtained by users in solving similar problems earlier.
In addition, the effectiveness of the proposed visualization tool is based on
the use of a deviation metaphor, which has shown good results in analyzing
multidimensional data that are similar in parameters to the test data. As a
visualization technique necessary to indicate missing values, excluding the
corresponding model nodes from the user's field of view and changing the
brightness of the image fragments of informative elements with missing values
are used (Fig. 8). At the same time, the simultaneous presence of a node name
and shaded element image fragment in the field of view is used as an active
pointer to the model area that requires the user's attention.
Fig. 8
Visualization of missing values. The name of each model node contains the
number of the informative element, as well as the value index in the record
In some
situations, for example, when the source data has high noise or a significant
number of missing values (this is typical for the analysis of data collected by
the CPS), the color-coding method becomes insufficient. This happens because
the user's attention becomes scattered during the process of interpreting a
heterogeneous saturated image. To reduce the burden on the user's perception, a
technique was used that reduces the visual model detailing while maintaining
its informativeness (Fig. 9). In the proposed solution, nodes of the visual
model with no values are shifted to a special image area (here, the central
point). For the user, the pointer to the area of focus no longer on nodes, but
rather on intersections between images of complete and damaged elements.
For solving the
problem of missing value recovery (when in a series of experiments, there is no
complete identity of the set of CPS measuring instruments or technical and
other failures occur during recording data), when using the considered metaphor
(Fig.10), it is necessary to select images of complete informative elements,
with most parameters close in values to images of elements with missing data.
To do this, it is enough to select the damaged element and determine number of
the informative element suitable for constructing a borrowing hypothesis. In
order to simplify the process of creating a borrowing hypothesis, the parameter
with corrupted data is shifted to the bottom part of model. This allows to
separate complete and damaged elements while maintaining a general
understanding of all values available for analysis. The proposed method is
particularly useful in situations where the source of direct borrowing is not
in the initial data and in order to form hypothesis, a full understanding of
the initial data and their features is necessary. This situation is the most
common for tasks of data processing collected by the CPS.
Fig.
9 Technique for reducing the detail of a model
Fig.
10 General view of the test data model
In this work, an
analysis of numerous factors that have a significant impact on the
effectiveness of using visual analytics tools for solving applied problems
associated with analyzing multidimensional data of different types was
performed. As a test example of such task, the problem of preparing and
researching multidimensional data collected by cyberphysical systems of
different types was examined. One of the features of such data is the lack of
some values in the descriptions of informative objects. A methodological
approach that relies on the advantages of visual perception in generating
hypotheses of interpretation and subsequent analysis has been proposed.
The possibility of
designing visualization tools for solving applied problems has been shown. This
includes the interactive determination of the most promising algorithm for
solving the problem of researching data of a specified type, as well as for
other applied tasks where achieving the research goal is complicated by the
lack of proven methods for utilizing the potential of modern information
technologies. The role of visualization and patterns of visual perception is
shown in both the process of researching initial data and controlling the
effectiveness of using the subjective potential of the user of visual analytics
tools.
In the task of
reconstructing the data collected by CPS, the metaphor of deviations can be
applied at the stage of evaluating data quality when using visual analytics
tools repeatedly. This is a typical situation for the analysis of data received
from monitoring systems, including those with a large number of different
sensors. Validation of the developed research methodology has demonstrated that
the user's understanding of expected values, formed in this case, makes the
deviation model a tool that speeds up the solution of analysis problem (by 25%,
according to experimental estimates). The metaphor of differences provides good
results in evaluating the possibility of borrowing values from similar
information objects and is therefore also recommended as a basis for designing
visualization tools at the stage of preliminary research. In addition, versions
of visual models that are based on generalization of perception and
interpretation of movement can increase the speed of solving research problems
by up to 40%.
This work was
supported by the Russian Science Foundation, grant No. 23-19-00342,
https://rscf.ru/en/project/23-19-00342/
1. V. Shklyar, A. A. Zakharova, è E. V. Vekhter, «Methods of solving problems of data analysis using analytical visual models», Sci. Vis., V. 9, Is. 4, P. 78–88, 2017, doi: 10.26583/sv.9.4.08.
2. Chen, «Science Mapping: A Systematic Review of the Literature», J. Data Inf. Sci., V. 2, Is. 2, P. 1–40, 2017, doi: 10.1515/jdis-2017-0006.
3. V. Shklyar, A. A. Zakharova, E. V. Vekhter, è A. J. Pak, «Visual modeling in an analysis of multidimensional data», J. Phys. Conf. Ser., V. 5, Is. 1, P. 125–128, 2018, doi: 10.1088/1742-6596/944/1/012127.
4. J. F. Rodrigues Jr., L. A. S. Romani, A. J. M. Traina, è C. Traina Jr., «Combining Visual Analytics and Content Based Data Retrieval Technology for Efficient Data Analysis», 4th International Conference Information Visualisation, IEEE, 2010, P. 61–67. doi: 10.1109/IV.2010.101.
5. A.E. Bondarev, V.A. Galaktionov. Generalized Computational Experiment and Visual Analysis of Multidimensional Data (2019). Scientific Visualization 11.4: 102 - 114, DOI: 10.26583/sv.11.4.09.
6. Vieira, P. Parsons, è V. Byrd, «Visual learning analytics of educational data: A systematic literature review and research agenda», Comput. Educ., 2018, doi: 10.1016/j.compedu.2018.03.018.
7. H. Chen è äð., «Uncertainty-Aware Multidimensional Ensemble Data Visualization and Exploration», IEEE Trans. Vis. Comput. Graph., V. 21, Is. 9, P. 1072–1086, 2015, doi: 10.1109/TVCG.2015.2410278.
8. V. L. Averbukh è äð., «Virtual reality as an instrument of computer visualization», â 2019 International Multi-Conference on Engineering, Computer and Information Sciences (SIBIRCON), 2019, P. 0786–0792. doi: 10.1109/SIBIRCON48586.2019.8957854.
9. M. A. Yalcin, N. Elmqvist, è B. B. Bederson, «Cognitive Stages in Visual Data Exploration», â Proceedings of the Beyond Time and Errors on Novel Evaluation Methods for Visualization - BELIV ’16, 2016. doi: 10.1145/2993901.2993902.
10. V. N. Kasyanov, Methods and tools for information visualization on the basis of attributed hierarchical graphs with ports, Siberian Aerospace Journal. 2023, Vol. 24, No. 1, P. 8–17. Doi: 10.31772/2712-8970-2023-24-1-8-17.
11. Batch è N. Elmqvist, «The Interactive Visualization Gap in Initial Exploratory Data Analysis», IEEE Trans. Vis. Comput. Graph., V. 24, Is. 1, P. 278–287, 2018, doi: 10.1109/TVCG.2017.2743990.
12. H. Guo, S. R. Gomez, C. Ziemkiewicz, è D. H. Laidlaw, «A Case Study Using Visualization Interaction Logs and Insight Metrics to Understand How Analysts Arrive at Insights», IEEE Trans. Vis. Comput. Graph., V. 22, Is. 1, P. 51–60, 2016, doi: 10.1109/TVCG.2015.2467613.
13. R. Pienta è äð., «VIGOR: Interactive Visual Exploration of Graph Query Results», IEEE Trans. Vis. Comput. Graph., ò. 24, âûï. 1, ññ. 215–225, ÿíâ. 2018, doi: 10.1109/TVCG.2017.2744898.
14. R. J. Crouser, L. Franklin, A. Endert, è K. Cook, «Toward Theoretical Techniques for Measuring the Use of Human Effort in Visual Analytic Systems», IEEE Trans. Vis. Comput. Graph., 2017, doi: 10.1109/TVCG.2016.2598460.
15. R. Akhmadeeva, Y. A. Zagorulko, è D. I. Mouromtsev, «Ontology-Based Information Extraction for Populating the Intelligent Scientific Internet Resources», â Knowledge Engineering and Semantic Web, A.-C. Ngonga Ngomo è P. Kremen, Ðåä., Cham: Springer International Publishing, 2016, P. 119–128.
16. V. Shklyar, A. À. Zakharova, E. V. Vekhter, è D. A. Zavyalov, «Visual detection of internal patterns in the empirical data», â Communications in Computer and Information Science, Volgograd, Russia: Springer Verlag, 2017, P. 215–230. doi: 10.1007/978-3-319-65551-2_16.
17. Chen è M. Song, «Visualizing a field of research: A methodology of systematic scientometric reviews», PloS One, V. 14, Is. 10, ñ. e0223994, 2019.
18. Ware, Foundation for a Science of Data Visualization. 2012. doi: 10.1016/B978-0-12-381464-7.00001-6.
19. Meirelles, «Diagramming às A Strategy For Solving Graphic Design Problems», Ïðîñìîòðåíî: 29 ôåâðàëü 2020 ã. [Îíëàéí]. Äîñòóïíî íà:
https://www.academia.edu/2749924/Diagramming_As_A_Strategy_For_Solving_Graphic_Design_Problems
20. S. Nison, Japanese Candlestick Charting Techniques. 2024.
21. J. H. Larkin è H. A. Simon, «Why a Diagram is (Sometimes) Worth Ten Thousand Words», Cogn. Sci., V. 11, Is. 1, P. 65–100, 1987, doi: 10.1111/j.1551-6708.1987.tb00863.x.
22. Branchini, U. Savardi, è I. Bianchi, «Productive thinking: Tlie role of perception and perceiving opposition», Gestalt Theory, V. 37, Is. 1, P. 7–24, 2015
23. Branchini, U. Savardi, è I. Bianchi, «Productive thinking: Tlie role of perception and perceiving opposition», Gestalt Theory, V. 37, Is. 1, P. 7–24, 2015
24. J. Choi, S. Jung, D. G. Park, J. Choo, è N. Elmqvist, «Visualizing for the Non-Visual: Enabling the Visually Impaired to Use Visualization», Comput. Graph. Forum, V. 38, Is. 3, P. 249–260, 2019, doi: 10.1111/cgf.13686.
