Business information systems nowadays should be thought of first of all as the decision-oriented systems supported by different types of subsystems. Multidimensional data visualization is an essential constituent of such systems, especially in the age of growing amounts of data to be interpreted and analyzed. As managers are faced with a federated environment and need to make timecritical decisions, data should be presented in a meaningful manner and easily understandable form. It is required more effective ways to cope with this situation. One of them is the visual presentation of complex data for human decisions. The paper focuses on the neural networks-based methods for visualization of big multidimensional datasets. The new strategy - to decrease the number of cycles of data reviews (passes of training data) up to the only one when training neural networks is proposed. The results of experiment on benchmark data to test this strategy are presented.
Characteristics of today’s world, such as globalization, dynamics and often unpredictable changes, huge amounts of data, are being observed on any of its entities. Even a philosophy in general beyond business information systems (BIS) is frequently changing. Nowadays, they should be thought of first of all as the decisionoriented systems supported by different types of subsystems. Multidimensional data visualization is an essential constituent of such systems because this approach enables to discover knowledge hidden in big datasets. As managers are faced with a federated environment and a need to make time-critical decisions, data should be presented in a meaningful manner and easily understandable form. However, when datasets are becoming increasingly large we require more effective ways to process, analyze and interpret these data.
We focus on neural networks-based methods for the visualization1 of big multidimensional datasets. Two unsupervised learning methods are considered: SAMANN (a feed-forward neural network to learn Sammon’s mapping) and SOM (Self-Organizing Map). To cope with data processing time problem the new strategy – to decrease the number of data passes (reviews) up to the only one when training neural networks is proposed. It is based on the assumption that huge amount of data includes many similar objects, so even in one pass, the neural network can see big amount of similar objects. After the training, network can be used for decision support – any number of new objects can be converted to meaningful form, i.e. presented as points on the plain.
Empirical research was carried out. To test the hypothesis, in other words, the proposed strategy, the experiment on a set of benchmark data was conducted. The additional evaluation of the outcome was done using the results of traditionally trained neural networks.
The structure of this paper is as follows. We start out, in Section 2, by positioning business information systems and their mission from today’s perspective. Section 3 considers data visualization – describes the strategy of training neural networks through a single pass of training data and its experimental investigation. Section 4 concludes the paper and comments on open issues. 2
The concept of information system2 has significantly changed throughout it’s more than 50 years history. These distinctions reflect its role and importance in a business enterprise and can be seen on the development approaches, methodologies, frameworks, architectural design decisions, technology.
At the very beginning, namely management information systems (MIS) served the business management. Their purpose was to cater to the information needs for planning, controlling and decision making. MIS is dependent on underlying transaction processing systems, but in fact, can itself be thought of as a transaction processing system, which possibly interacts with a decision-support subsystem. From technological point of view it was a set of applications centered around a database.
This philosophy has changed at the very beginning of the 21th century when it was
realized that an enterprise system should be developed as a whole [
Nowadays, information system should be thought of first of all as the
decisionoriented system supported by different types of subsystems. The previous concept of
IS cannot dominate, as manufacturing enterprises as a rule focuses on their core
competences, so unable to produce alone. Decision-making is federated and
synchronized between different divisions, within or between enterprises, to achieve
total and autonomous optimization [
One of the challengers in this context is the ability to process big amounts of data in near-real time to make the decisions. 3
3.1
Big multidimensional data brings new challenges to data analysis because large volumes and different varieties must be taken into account. In many cases, data is just being generated faster than it can be analyzed. To analyze big data, many data mining and machine learning algorithms have been developed. We focus on dimensionality reduction algorithms which reduce data dimensionality from original high dimension space to target dimension (2D in visualization case). Data visualization is the presentation of multidimensional data in some graphical form. As more and more data should be collected and analyzed, it is very important to see analytical results presented visually, find dependences among a lot of objects.
Visualization is one of the basic operations in the toolbox of data analysts. Given a large set of some measured variables, the main idea is to represent them with a reduced set of more informative variables. Another reason for reducing the dimensionality is to reduce computational load in further processing. Today's large multidimensional datasets contain huge amount of data that becoming almost impossible to manually analyze them to extract valuable information. We require more effective ways to display, analyze and interpret the information contained within them.
Data from the real world are frequently described by an array of features !!,!!, … , !!. Any feature may take some numerical values. A combination of values of all features characterizes a particular data object !! = (!!!,!!!, … , !!"), ! ∈ 1, … , ! from the whole set !!, !!, … , !!, where n is the number of features, m is the number of analyzed objects. If !!, !!, … , !! are described by more than one feature, the data are called multidimensional data. Often the objects are interpreted as points in the n-dimensional space !!, where n defines the dimensionality of the space. In fact, we have a table of numerical data !!!, ! = 1, … , !, ! = 1, … , ! for the analysis. An intuitive idea is to present multidimensional data, stored in such a table, in some visual form. It is a complicated problem considered by many researchers, as solution allows the human to gain a deeper insight into the data, draw conclusions, and directly interact with the data. A type of multidimensional data visualization is based on dimensionality reduction. The goal of dimensionality reduction is to represent the input data in a lower-dimensional space so that certain properties (e. g., clusters, outliers) of the structure of this dataset were preserved as faithfully as possible. 3.2
A comprehensive review of the dimensionality reduction methods is presented in [
The effectiveness of PCA is limited by its global linearity. The MDS method is nonlinear method, however, unsuitable for large datasets: it requires too much computational resources. Therefore, the combinations of different data visualization methods are under active development today. The combination of different methods can be applied to make more efficient data analysis, while minimizing the shortcomings of individual methods.
Artificial neural networks (ANNs) may also be used for dimensionality reduction
and data visualization. The MDS got some attention from the neural network
researchers [11], [12]. As a result, several neural networks based methods for the
visualization of big multidimensional datasets have been proposed, including
SAMANN [
The most ANN based visualization methods are unsuitable for large datasets due to the demand of huge computational resources. One possible solution employs the hardware – increased memory, parallel processing and grid computing. The second solution is to go the other way and to develop more mature neural networks based visualization theory. Therefore the new strategies, approaches and methods for training artificial neural networks are required. 3.2.1
A particular case of the metric MDS method is Sammon’s mapping. It tries to
optimize a projection error that describes how well the pairwise distances in a dataset
are preserved. The application of original Sammon’s mapping is not suitable for large
datasets. Another disadvantage of this method is that the whole mapping procedure
has to be repeated when a new data point has to be mapped. The back
propagationlike learning rule (called SAMANN rule) [
The self-organizing map (SOM) is a class of neural networks that are trained in an
unsupervised way using a competitive learning [
To visualize big multidimensional data using SAMANN and SOM a new strategy for training these networks is proposed. The advantage of this strategy is that the network can be trained to visualize the multidimensional data through a single pass of training data. After the training, the network can be used for visual presentation of the desirable number of multidimensional objects on the plain. The strategy is based on the assumption that huge amount of data includes many similar objects, so even during one pass, the neural network can see big amount of similar objects.
A new strategy of big multidimensional data visualization using SAMANN is presented in Fig. 1: (1) training of SAMANN neural network using single pass (only 1 iteration); calculating its weights; (2) graphical presentation (visualization) of the dataset; (3) graphical presentation of new previously unseen points using calculated weights without additional neural network training.
A new strategy using SOM is presented in Fig. 2: (1) training of the SOM neural network using single pass; SOM winning neurons are calculated; (2) visualization of two-dimensional points that are two-dimensional representations of the codebook vectors of the winning neurons by MDS; (3) graphical presentation of the dataset; (4) graphical presentation of new previously unseen objects using the winning neurons by MDS without additional SOM training.
1
single pass of training data;
2
3
Ellipsoid dataset have been used to investigate the ability to visualize big multidimensional dataset using SAMANN and SOM. The ellipsoidal dataset consists of 7354 10-dimensional points from 10 overlapping ellipsoidal-type clusters. In the experiments, dataset, obtained using the ellipsoidal cluster generator [14], is used. The results of the experiments of multidimensional data visualization through a single pass of training data are presented in Fig. 3a and 4a. The points of the dataset are marked by black triangles. The circles correspond to the new points that were not used for training. For additional validation, the results of traditionally trained neural networks are presented in Fig. 3b and 4b.
The comparison of the results using SAMANN and SOM shows that it is possible to get the suitable projections of the primary dataset using single pass of training data and to visualize new points. The experiments show that even after one pass rather appropriate projection can be obtained. a) a) 1" 0.8" 0.6" 0.4" 0.2" 0" 1.5" 1" 0.5" )1" )1.5" 0" 0.2" 0.4" 0.6" 0.8" 1" 0" 0.2" 0.4" 0.6" 0.8" 1" Multidimensional data visualization is an essential constituent of business information systems, especially in the age of growing amounts of data to be interpreted and analyzed. Therefore, it is required the more mature neural networks-based visualization theory.
The new strategy to decrease the passes of training data up to the only one when training neural networks is proposed and examined. The results of experiment on the benchmark data to test this strategy allow us to conclude that the unsupervised learning of SAMANN and SOM neural networks are effective in producing the visual projections of the big multidimensional data, where we do not need any additional knowledge on the objects – the known numerical values of the features are sufficient. The obtained visualization results are good and computational expenses are acceptable if compare with traditional learning when a lot of iterations are required.
Further research should be focused on the theoretical background of such single pass strategy as well as on the discovering new domains (e. g., streaming data analysis [15]) where big multidimensional data are required to be visualized when reaching proper human decisions.
This work has been supported by the project “Theoretical and Engineering Aspects of e-Service Technology Development and Application in High-Performance Computing Platforms” (No. VP1-3.1-ŠMM-08-K-01-010) funded by the European Social Fund.