IJCSNS International Journal of Computer Science and Network Security, VOL.11 No.
6, June 2011 243
Prediction of Students' Educational Status Using CART
Algorithm, Neural Network, and Increase in Prediction Precision
Using Combinational Model
Mohammad Tari, Behrouz Minaei , Ahmad Farahi, Mohammad Niknam Pirzadeh
Tehran Payame Noor Iran University of Science Payame Noor Tehran Payame Noor
University and Technology University University
Future prediction in various fields has ever been
Summary interesting and attractive for human being. Surely, it can
In this paper, using CART and Neural Network data be said that future prediction and transformations trend in
mining algorithms we have dealt with the study of all areas is one of the basic and constant distresses of high
effective factors in the rate of becoming unqualified and intermediate-level managers. However, there have
among students of Payame Noor University of Qom always been numerous difficulties against it that have
Province. After creating models using algorithms made precise and reliable predictions almost impossible.
separately, we have combined models in Clementine 12.0 Existence of too many and in many cases hidden
software through Ensemble Node and have created the parameters have changed such instances into highly
combinational model and finally with Analysis Node we complicated problems that mathematical gigantic
have evaluated models and compared results. In the algorithms have become incapable of presenting an
created model using neural network algorithm, in input appropriate technique to build an efficient prediction
layer 44 neurons, in hidden layer 3 neurons, and in output model.
layer 1 neuron are created, also in CART model a decision In two recent decades, simultaneous with emergence of
tree with depth of 5 is created. Considering that the artificial intelligence and its combination with well-
number of fields in data bank is high, using Feature established science of statistics along with advanced and
Selection Node and selecting the target, we have deleted innovative algorithms like genetic algorithm, metaheuristic
those fields that had less influence on the target. This methods, artificial neural networks, and … a wide
matter has decreased the model complexity to some extent. revolution has been made in this arena.
Key words: Data mining as a new and flourishing area in the field of
mining, neural network, CART, Payame Noor University of Qom presenting predictor models have applied and combined
Province, Ensemble different kinds of statistical techniques, artificial intelligent
and innovative algorithms so that it can discover reliable
algorithms and models from among of gigantic data
1. Introduction storages to predict intended parameters.
that education always encounters too much data and
information about universities, students, faculty members, 2. CRISP Methodology
personnel, pecuniary resources, etc and most of time, this
data can contain valuable information and patterns, so it Data mining1 techniques are among scientific modern
seems that one of the most important applications of data techniques that are used in description, explanation,
mining is in higher education. Nowadays, there are wide prediction, and control of phenomena. These techniques
data banks of students' characteristics that include family, measure, explain, and predict the correlation degree among
educational, and … characteristics. Finding patterns and variables. Data mining methods influence not only on
knowledge concealed in this information can considerably analytic aspects of studies but also on designing and tools
help decision-makers of higher education arena. of data collection for decision-making and solving
Finding concealed patterns and knowledge from problems. The most successful data mining projects are
educational systems' data can considerably help decision- implemented in the framework of a standard process,
makers of higher education arena in advancement and which is presented by a work team in SPPS Company in
improvement of educational processes like planning, the form of projects named CRISP-DM. According to
enrollment, evaluation, and consultation. CRISP-DM, a certain data-mining project has a six-stage
life cycle that shows succession of stages. Each stage of
Manuscript received June 5, 2011
Manuscript revised June 20, 2011
244 IJCSNS International Journal of Computer Science and Network Security, VOL.11 No.6, June 2011
stages' succession often includes the result of correlation 4.1. Tables Combination
of previous stages. The most important correlation among
stages is showing arrows. Recurrent feature of CRISP In this stage that is a part of data preparation, we combine
indicates external cycle that often leads to a solution for tables and the result is a new table with the specification of
research or business problem with additional remarkable table 3 that is used as data mine for algorithms.
questions. In figure 1, you can observe stages of this Through nodes that are data preparation special, we
methodology [1]. implement some changes on this data mine.
Table (3): List of created combinational table fields
Field Field type Field name Field type
name
Gender Male/female unqualified numerical
Native Yes/no successively numerical
unqualified
Field Fields of study Passed numerical
of list credits
study
Course Formal/compreh Kind of ration
ensive admission
Total numerical age numerical
averag
Figure (1): Stages of CRISP-DM Methodology
e
3. Table Applied 4.2. Data anomaly3 recognition
Table (1): Basic information One of the methods of preparing data is to find
Student First Kind of Admission Birth native anomalous data from the tables and delete them in order to
Number name and admission Term Date gain results that are more precise through models. To do
surname in this, we attach data mine to Anomaly node and next, we
university
add the created model to the project, and then by
Some unnecessary fields have been deleted.
adjustment of model, anomalous data in models creation
Table (2): Lessons
Lesso Lesso Ter Sco Theor Practical Explanations
are deleted. In the figure, you observe the way of model
n's n's m re etical credit creation.
code name credit
4. Data preparation
Pre processing2: The importance of data preparation is
because of the reality that "lack of data of quality is equal
to lack of quality in results of mining" and improper input
leads to improper output. In table 2, the importance of data
preparation is compared to the importance of other steps of
knowledge discovery using data mining. [2]
Figure (2): Data Anomaly Model
Data mining step Percent of Percent of As you can see in figure number 2, first, we attached data
time spent importance in mine to Anomaly algorithm and after creating the model,
from total work's final we added it to the project. Indeed, the technique of
work success anomalous data recognition is use of clustering methods.
Data preparation 75 75
Data investigation 20 15 4.3. Feature Selection Technique
Data modeling 5 10
A technique is used to decrease the number of variables
Table (2): comparison of data preparation importance before implementation of data mining algorithm. Irrelevant
IJCSNS International Journal of Computer Science and Network Security, VOL.11 No.6, June 2011 245
variables can have negative effect on prediction duty, or processor system that a large number of them in the form
can complicate the calculations. of a complex work together like brain hormones in order
In order to implement algorithms on data banks, we to solve special problems like pattern detection or data
decreased the number of required fields through this classification through learning process. [3]
technique.
To do this work, first, we create a mine node (figure 3) and
then link data mine, which Anomaly technique has been 6. Clementine software
implemented on it to that feature selection.
First, Ulshan, Friedman, Barmyman, and Stone designed
CART algorithm for trees of regression and classification
in 1984. The operation method of this algorithm is named
Surrogate Spiliting. This algorithm includes a recurrent
method. In every stage, CART algorithm divides
instructional records into two subsets so that the records of
each subset are more homogeneous than previous subsets.
These divisions continue until conditions of stop are
established. In CART, the best breaking point or
assignment of the value of impurity parameter is
determined. If the best break for a branch makes impurity
F
less than the defined extent, that split will not be made.
igure (3): implementation of Feature Selection model
Here, the concept of impurity is assigned to the degree of
similarity between target field value and records arrived to
For this model, successively unqualified field was selected
a node. If 100 percent of the samples in a node are placed
as target field and after algorithm implementation that 28
in a specific category of the target field, that node is
fields have been selected as inputs, 3 fields were deleted
named pure. It is remarkable that in CART algorithm, a
from list of fields and 25 fields were selected as inputs of
foreteller field may be often used in different levels of
neural network and CART algorithms.
decision tree. In addition, this algorithm supports
This method is one of the supervisory learning techniques.
categorical and continues types of foreteller and target
When outcomes are two classes, linear regression is not so
fields. [3]
efficient; in this state, applying this technique is more
appropriate. The other point is that this method is a
nonlinear regression technique and it is not necessary for 7. Model Creation
data to have linear state. If we want to say the reason for
use of logistic regression, we should argue that in linear Because of lengthiness of data preparation part, we have
regression not only outcomes should be in numerical given up to cite its stages in this section.
forms, but also variables should be in numerical forms as To create the model, first we add a data mine node to the
well. Hence, the states that are in sort form should change project and then add a partition node to the project in order
to numerical forms. to divide data to two parts of test and learning that 30
percent of data are test data and its 70 percent are
4.4. Filtration and deleting wide data instructional data.
Next stage after data conversion is data type determination
To make a better model, in addition to fields deleted by using type node; after this stage, feature selection
Feature Selection algorithm, using filter node we delete capability is used for determination of fields' importance in
those fields that do not influence on models' output. prediction. This leads to the selection of those fields that
Therefore, from 28 total fields, 8 fields were deleted, and have more importance in prediction.
finally 15 fields were transferred to algorithms for model In subsequent stage, we add models, which are CART and
implementation. Neural Net to the project.
In addition, using Select node, we deleted from the table When models are attached, we implement them and then
those data, which their values were null. link models' outputs to each other; and after linking
models, using Ensemble node we combine them with each
other. In figure 4, you observe the created model.
5. Neural Network4
Neural Network or ANN is an instance of processing
system, which is inspired from biological neural
networks like brain. Key organ of this new structure is data
246 IJCSNS International Journal of Computer Science and Network Security, VOL.11 No.6, June 2011
Figure (5): Neural Network's Gain Graph
Figure (4): the created combinational model
7.1. Evaluation of Neural Network Model
To evaluate neural network model, first we add an analysis
node to the project and then attach it to the created model
and implement it.
Minimum Error -1.68
Maximum Error 2.722
Mean Error -0.045 Figure (6): the importance of variables in neural network
Mean Absolute Error 0.28 model
Standard Deviation 0.35
Linear correlation 0.987 7.2. Evaluation of CART Model
Occurrences 1,678
In order to evaluate CART model, first we add an analysis
node to the created model and implement it.
Minimum Error: the least difference between predicted
Minimum Error -3.531
data and observed data
Maximum Error 4.469
Maximum Error: the most difference between predicted
Mean Error -0.0
data and observed data
Mean Absolute Error 0.276
Mean Error: the observance of errors' average among the
Standard Deviation 0.516
total of records
Linear correlation 0.971
Mean Absolute Error: the observance of the average of
Occurrences 1,678
errors' absolute value among the total of records
Standard Deviation: observance of Standard deviation of
the total of errors
Linear Correlation: linear Correlation between observed
data and real data
Occurrences: the number of records used in prediction
In figure 6, the degree of correlation or the importance of
fields in target field has been shown that fields of
successively unqualified and number of unqualified are the
most important and field of term type is the least important
one in target field prediction.
Figure (7): CART Model's Gain Graph
IJCSNS International Journal of Computer Science and Network Security, VOL.11 No.6, June 2011 247
Minimum Error -1.818
Maximum Error 3.596
Mean Error -0.022
Mean Absolute Error 0.24
Standard Deviation 0.348
Linear correlation 0.992
Occurrences 1,678
As we see, the degree of linear correlation has increased to
0.992; that the precision of model has increased eminently
and model's errors have decreased.
Figure (8): the importance of variables in CART Model
In this stage, using table No.1 that is a table with almost
one million records for 5 successive years log of students'
use of internet, we can find maximum, minimum, and
average values in terms of kilobytes for students' monthly
use. Figure (10): Combinational model's Gain Graph
7.3. Evaluation of Combinational Model
8. Conclusion
Ensemble node combines 2 or more models so that a more
precision is gained compared to models that are Contemporary universities and higher education
implemented separately; through combining predictions institutions are drowned in a mass of data and information
from prediction models, limitations that are created in that in most cases, their use is limited to current affairs and
single models are deleted and a more precision can be data are not yet used in strategic decision-making. Data
achieved. Most of time combinational models have a better mining which its use is developing day-to-day can lead to
result than separate models. After using Ensemble node, the use of existing information in higher education
we can use Analysis node to compare the results of institutions and centers in strategic decision-making
combinational model to separate model. domains. In addition, created models from instructional
data, using data mining algorithms can be utilized as a
decision support system in educational systems and play
an important role in advancement of universities' scientific
level.
REFERENCES
[1] http://www.crisp-dm.org
[2] Andrew W. Moore. "Regression and Classification with
Neural Networks". School of Computer Science Carnegie
Mellon University. 2001.
[3] Mosavi, M.R., "A Comparative Study between Performance
Figure(9):Models' combination through ensemble node of Recurrent Neural Network and Kalman Filter for DGPS
Corrections Prediction", IEEE Conference on Signal
As you see in figure number 9, after creation of 2 models Processing (ICSP 2004), China, Vol.1, pp.356-359, August
31-4, September, 2004.
we link them to each other and then combine them through
[4] Ramos, V. and Abraham, A., "Evolving a Stigmergic Self-
Ensemble node. After implementation of combinational Organized Data-Mining", Conf. on Intelligent Systems,
model, in order to evaluate model, we attach it to Analysis Design and Applications (ISDA-04), 2004, pp.725-730.
mode.
