EXTRACTING USEFUL RULES THROUGH IMPROVED DECISION TREE INDUCTION USING INFORMATION ENTROPY

International Journal of Information Sciences and Techniques (IJIST) Vol.3, No.1, January 2013
DOI : 10.5121/ijist.2013.3103 27
EXTRACTING USEFUL RULES THROUGH
IMPROVED DECISION TREE INDUCTION USING
INFORMATION ENTROPY
Mohd. Mahmood Ali1
, Mohd. S. Qaseem2
, Lakshmi Rajamani3
, A. Govardhan4
1
Dept. of C.S.E., Muffakhamjah College of Engg. & Tech., Hyderabad, India
mahmoodedu@gmail.com
2
Dept. of C.S.E., Nizam Institute of Engg. & Tech., Hyderabad, India
ms_qaseem@yahoo.com
3
Dept. of C.S.E., Univ. College of Engg., O.U., Hyderabad, India
drlakshmiraja@gmail.com
4
Dept. of C.S.E., JNTUH., Hyderabad, India
dejntuh@jntuh.ac.in
ABSTRACT
Classification is widely used technique in the data mining domain, where scalability and efficiency are the
immediate problems in classification algorithms for large databases. We suggest improvements to the
existing C4.5 decision tree algorithm. In this paper attribute oriented induction (AOI) and relevance
analysis are incorporated with concept hierarchy’s knowledge and HeightBalancePriority algorithm for
construction of decision tree along with Multi level mining. The assignment of priorities to attributes is
done by evaluating information entropy, at different levels of abstraction for building decision tree using
HeightBalancePriority algorithm. Modified DMQL queries are used to understand and explore the
shortcomings of the decision trees generated by C4.5 classifier for education dataset and the results are
compared with the proposed approach.
KEYWORDS
Attribute oriented induction (AOI), Concept hierarchy, Data Mining Query Language (DMQL),
HeightBalancePriority algorithm, Information entropy, C4.5 classifier.
1. INTRODUCTION
Classification is a supervised learning technique in data mining where training data is given to
classifier that builds classification rules. Later if test data, is given to classifier, it will predict the
values for unknown classes.
C4.5 classifier [1],[2], a well-liked tree based classifier, is used to generate decision tree from a
set of training examples. Nowadays C4.5 is renamed as J48 classifier in WEKA tool, an open
source data mining tool. The heuristic function used in this classifier is based on the concept of
information entropy.
Induction of decision trees from very large training sets has been previously addressed by SLIQ
[3] and SPRINT [4], but the data stored is without generalization. The generalization concept for
evaluating classification rules using DMQL in data cube is proposed [5]. Decision tree

28
construction process performed on very large datasets leads to bushy or meaningless
results. This issue which is addressed in data generalization and summarization based
characterization [5], consist of three steps attribute-oriented induction, where the low-level
data is generalized to high-level data using the concept hierarchies, relevance analysis [6], and
multi-level mining, where decision trees can be induced at different levels of abstraction.
The integration of these steps leads to efficient, high quality and the elegant handling of
continuous and noisy data. An inherent weakness of C4.5 classifier, is that the information
gain attribute selection criterion favors many-valued attributes, due to which some of the
attributes are pruned because of less information gain in decision tree construction. Ultimately
this leads to less classification rules. C4.5 classifier encounters the over-branching problem
caused by unnecessary partitioning of the data which is of least importance to users. Therefore,
we propose a HeightBalancePriority algorithm which constitutes NodeMerge and
HeightBalance algorithms [11] that allows merging of nodes in the decision tree thereby,
discouraging over-partitioning of the data. This algorithm also uses the concept of Height
Balancing in the decision tree based on priority checks at every node based on attribute values
selected for different attributes using information entropy. This enhances the overall
performance, as the final decision tree constructed is efficient enough to derive the
complete classification rules effectively avoiding over-branching problem of unnecessary
attributes.
In this paper, we emphasized on decision tree construction and classification rules derived
based on priority by using information entropy, which gives importance to selected attributes
that was neglected by C4.5 classifier. In the Remainder of the paper, for clarity and better
understanding, the Modified DMQL queries are used that explore the dataset used for decision
tree construction.
The paper is organized as follows. Section 2, describes the data generalization and
summarization based characterization. Section 3 briefly, explains about the proposed
algorithms used for decision tree construction. The Decision tree formed using C4.5, yet
suffers from deficiencies. To overcome the priority based decision tree construction method is
discussed in Section 4. In Section 5, some more examples are taken to explore C4.5,
deficiencies and results obtained are compared with proposed approach. Section 6
summarizes the methodology for priority based Decision tree construction and future
challenges.
2. CLASSIFICATION USING DECISION TREE INDUCTION
Scalability and efficiency issues for classification techniques are addressed for large databases
that has improved C4.5 classifier, using the following four steps:
• Generalization by AOI, which compresses training data. This includes storage of
generalized data in data cube to allow fast accessing [8].
• Relevance analysis, that removes irrelevant attributes, thereby, further compacting
training data.
• Multi-level mining, which combines the induction of decision trees with
knowledge in concept hierarchies.
• Priority based height balance trees using Entropy.
The top three methodologies are discussed in Generalization and Decision tree induction for
efficient classification [7], we have also used the above three methodologies with slight
improvements, where as the priority based decision tree construction using Information
entropy, is discussed in Section 4.

29
2.1. Generalization by AOI (Attribute oriented Induction)
AOI [9], a knowledge discovery tool which allows the generalization of data, offers two major
advantages for the mining of large databases. Firstly, it allows the raw data to be handled at higher
conceptual levels. Generalization is performed with the use of attribute concept hierarchies, where
the leaves of a given attribute’s concept hierarchy correspond to the attribute’s values in the data
(referred to as primitive level data) [10]. Secondly, generalization of the training data is achieved
by replacing primitive level data by higher level concepts. Hence, attribute- oriented induction
allows the user to view the data at more meaningful abstractions.
Furthermore, AOI, addresses the scalability issue by compressing the training data. The generalized
training data will be much more compact than the original training set, and hence, will involve
fewer input/output operations. With the help of AOI, many-valued attributes in the selection of
determinant attributes are avoided since AOI reduces large number of attribute values to small set
of distinct values according to the specified thresholds.
AOI also performs generalization by attribute removal [9]. In this technique, an attribute having a
large number of distinct values is removed if there is no higher level concept for it. Attribute
removal further compact the training data and reduces the bushiness of resulting trees. Hence, aside
from increasing efficiency, AOI may result in classification trees that are more understandable,
smaller, and therefore easier to interpret than trees obtained from methods operating on non-
generalized (larger) sets of low-level data. The degree of generalization is controlled by an
empirically set generalization threshold. If the number of distinct values of an attribute is less
than or equal to this threshold, then further generalization of the attribute is halted. We consider
a simple example to explain all the detail steps to generalize the final classification tree and
find out the classification rules. Table 1, depicts a raw training data of class of average education
level is used, region wise, with the count in relation with the country part to which that place
belongs around the world.
Table 1. Attributes with four attribute values.
Average Education level Region Family Income per year Income Level
Illiterate Cuba.north $ 899 Low
fouryearscollege USA.east $ 30000 Medium
Fouryearscollege USA.south $ 38000 High
Fouryearscollege USA.middle $ 32000 High
twoyearscollege USA.middle $ 30400 High
Graduate school China.south $ 38000 High
Elementary school Cuba.north $ 990 Low
High school India.east $ 7839 Low
Fouryearscollege China.east $ 30000 Medium
Graduate school China.west $ 38000 High
Junior High China.south $ 3800 Low
Twoyearscollege India.south $ 20000 Medium
Fouryearscollege USA.west $ 20000 Medium
Graduate school China.west $ 38000 High
Ph. D India.south $ 50000 High
The Generalization using AOI for the attribute WORLD is depicted in Figure 1. Similarly
the concept hierarchies for age, income, and education can also be represented. We illustrate
ideas of attribute-oriented induction (AOI) with an example for edu_dataset shown in table 1,

30
generalized using the concept hierarchy. AOI is performed on set of relevant data. An
intermediate relation, achieved by concept ascension using concept hierarchies as shown
in Figure 1, attribute region normalized to country as shown in Table 2. Owing to attribute
removal technique, attribute values INDIA.EAST, INDIA.WEST, INDIA.SOUTH, are replaced
with INDIA, as depicted in Figure 1.
{CUBA,CHINA,USA,INDIA} {WORLD};
{CUBA.NORTH, CUBA.SOUTH} {CUBA};
{USA.EAST,USA.WEST,USA.SOUTH,USA.MIDDLE,} {USA};
{CHINA.SOUTH,CHINA.EAST,CHINA.WEST} {CHINA};
{INDIA.EAST,INDIA.SOUTH,INDIA.WEST} { INDIA};
Figure 1. Concept hierarchy for attribute values of region
Identical tuples for Table 1 merged while collecting the count information shown in Table 2,
which is achieved by applying modified data mining query language (DMQL) queries as
shown with example 2.1. The resultant table obtained is substantially smaller than the
original table, generalized with task-relevant data as shown in Table 2.
Example 2.1. Classification task.
Classify Region_Data till World_Data replace {USA,Cuba,India,China} attribute_values
with new_attribute Country
in relevance to avg_edu_level,country new_attribute count
from edu_dataset
where Country_Data={”India“,”USA”,”Cuba”,”China”}
and Region_Data={”India.east”,”India.west”,”India.south”,
”USA.east”,”USA.west”,”USA.south”,” USA.middle”,”Cuba.north”,
”Cuba.south”, ”China.south”,”China.east”,”China.west”}
In the above modified DMQL query the bold words represent reserve keywords. The keyword
Classify constitutes the proposed approach algorithms, which are discussed in this paper.
Table 2. Generalized data set obtained after AOI using DMQL query of example 2.1 with count
as new attribute.
Average Education level Country Family Income per year Count
Illiterate Cuba $1899 2
FourYearsCollege USA $120,000 4
GraduateSchool China $114,000 3
TwoYearsCollege India $40,000 2
FourYearsCollege China $30000 1
2.2. Relevance Analysis
The uncertainty coefficient U(A) for attribute A is used to further reduce the size of the generalized
training data as shown in Equation (1). U(A) is obtained by normalizing the information gain of A
so that U(A) ranges from 0 (meaning statistical independence between A and the classifying
attribute) to 1 (strongest degree of relevance between the two attributes). The user has the option
of retaining either the n most relevant attributes or all attributes whose uncertainty coefficient
value is greater than a pre-specified uncertainty threshold, where n and the threshold are user-

31
1 2
p
defined. Note that it is much more efficient to apply the relevance analysis to the generalized data
rather than to the original training data.
U(A)= I(p1, p2,…,pm,) – E(A) Equation... (1)
I(p1, p2,…, pm,)
Where: I(p1 , ,…, pm,) = - ∑m pi/ p log2 pi /p
E(A) = ∑ p1j,+ …+ pmj I(p1j,…, pm,) Equation… (2)
j=1
p
Gain(A)= I(p1, p2,…, pm,) – E(A) Equation… (3)
Here P is the set of the final generalized training data, where P contains m distinct values
defining with the output distinct output class Pi (for i = 1, 2, 3,…,m) and P contains pi samples
for each Pi, then the expected information i.e. Information Entropy needed to classify a
given sample is I(p1, p2,…, pm ). For example: we have the attribute A with the generalized final
value {a1 ,a2 ,a3 , ...,ak} can be partition P into {C1 ,C2, C3 , ...,Ck} , where Cj contain
those samples in C that have value aj of A. The expected information based on partitioning by
A is given by E(A) Equation (2), which is the average of the expected information. The Gain(A)
given in Equation (3), is the difference of the two calculations. If the uncertainty coefficient for
attribute A is 0, which means no matter how we partition the attribute A, we can get nothing
lose information. So the attributes A have no effect on the building of the final decision tree. If
U(A) is 1, mean that we can use this attribute to classify the final decision tree. This is similar
to find the max goodness in the class to find which attribute we can use to classify the final
decision tree. After the relevance analysis, we can get rid of some attribute and further compact
the training data based on selection of attributes by users.
Entropy for income class which has attribute values (High, Medium, Low):
I(s1,s2,s3)=I(5,5,7)=-
= [-0.29411764705882 * log2(0.29411764705882)-
0.29411764705882 * log2(0.29411764705882)-
0.41176470588235 * log2(0.41176470588235)]
= 1.56565311164580141944
If we want to choose one of the attribute as root node then, we have to calculate Information
Entropy for all other attributes present, in the education dataset, in most of cases the continuous
values are given preference to be the root node, if not multi-valued attribute values of an attribute is
selected as root node. This depends effectively on the highest information gain for the attribute
values present in the attributes (Average education level, Region, family income per year). Let us
calculate the information gain for all other attributes using equation (1) and equation (2) discussed
above.
We have to still compute the expected information for each of these distributions
For Avg education level = “Illiterate”
s11=2 ; s21=0; & s31=0
p

32
I(s11,s21,s31)= = 0
For Avg education level =”4yearscollege”
s12=2; s22 =3; &s32=0
I(s12,s22,s32)=
For Avg education level =”2yearscollege”
s13=1; s23 =2; & s33=0
I(s13,s23,s33)=
For Avg education level =”Graduate school”
s14=3; s24 =0; &s33=0
I(s14,s24,s34)=
For Avg education level =”Elementary school”
s15=1; s25 =0; & s35=0
I(s15,s25,s35)=
Similarly entropy for attribute values of attribute Avg education level = ”Ph.D”, “High School",
”Junior High” is 0; Using the equation (2) given above , if the samples are partitioned
according to Avg education level, then
E(Avg education level) =
= 0.44762470588235
Hence, the gain in information from such a partitioning would be calculated using equation (3)
given above:
Gain (Avg education level) = I(s1,s2,s3) E(Avg education level)
=1.56565311164580141944 0.44762470588235
= 1.11802840576345141944
Similarly, we have calculated Gain(Region)=1.3; Gain(family income per year)=0; We can say
that Gain is high for Region, so we have to select Region as root node. After the decision tree is
built, tree pruning phase is applied, resulting in removing of the nodes(attributes) which is of
user interest from the tree. Cross-validation technique analyzes the final constructed tree that
estimate errors, and rebuilds the decision tree as used in WEKA tool discussed in Section 4 and
Section 5. Some of the classification rules are also skipped resulting in inadequate decision trees.
So to avoid the above problems, in our approach, we propose algorithms discussed below which
give priority preference to the nodes (attributes) for decision tree construction with hidden
classification rules in Section 3, and with example in Section 4. The DMQL queries are used in
this paper to understand the propose work clearly, that integrate our propose algorithms.

33
2.3. Multi-level mining
The third step of decision tree construction method is multilevel mining. This combines decision
tree induction of the generalized data obtained in Sections 2.1 and 2.2 (Attribute- oriented
induction and relevance analysis) with knowledge in the concept hierarchies. The induction of
decision trees is done at different levels of abstraction by employing the knowledge stored in the
concept hierarchies. Furthermore, once a decision tree has been derived [5], the concept
hierarchies can be used to generalize individual nodes in the tree, allowing attribute roll-up or
drill- down, and reclassification of the data for the newly specified abstraction level. This
depends on information entropy and the selection of attributes by the user.
The main idea is to construct a decision tree based on these proposed steps and prune it
accordingly based on priority using Height Balancing tree concept [14], without losing any of
the classification rules. The Decision Tree Construction Algorithm 1 is shown in Figure 2,
which constructs a decision tree for the given training data. The stopping criteria, depends not only
on maximum information entropy but also selection of the attributes by the users and its priority.
In some cases if the attributes selected by the user are not in the data set then thresholds
plays a vital role as previously addressed [7] for decision tree construction. Apart from
generalization threshold, we used two other thresholds for improving the efficiency namely,
exception threshold (€) and classification threshold (ĸ). Because of the recursive partitioning,
some resulting data subsets may become so small that partitioning them further would have no
statistically significant basis. These “insignificant” data subsets are statistically determined by
the exception threshold. If the portion of samples in a given subset is less than the threshold,
further partitioning of the subset is halted. Instead, a leaf node is created which stores the subset
and class distribution of the subset samples.
The splitting-criterion in the Algorithm 1, deals with both the threshold constraints and also
information gain calculation for the data. In this process, the candidate with maximum
information gain is selected as “test” attribute and is partitioned. The condition, if the
frequency of the majority class in a given subset is greater than the classification threshold, or if
the percentage of training objects represented by the subset is less than the exception
threshold, is used to terminate classification otherwise further classification will be performed
recursively until all the attributes are not selected which is of users interest. The algorithm operates
on training data that has been generalized to an intermediate level by AOI, and for which
unimportant attributes have been removed by relevance analysis, if attributes selected are not
present in the users list.
Algorithm 1. Decision Tree Construction
Input: Pre-processed Dataset without noise.
Output: Final well-balanced Decision tree built based on priority attributes selected by users
interest.
DecisionTree (Node n, DataPartition D)
{ Apply AOI-Method to D to find splitting-criterion of node n
Let k be the number of children of n if k>O do
Create k children c1, c2,..., ck of n
Use splitting-criterion to partition D into D1, D2..., Dk
for i = 1 to k do
DecisionTree(ci, Di)
end for
endif
//Assign priority to the nodes based on the attribute

34
Call HeightBalancePriority(HeightBalance, Priority_Attribute)
//Calls Algorithm 4
ALLOCATE highest information entropy attributes to
PRIORITY attribute/
}
Figure 2. Decision tree construction algorithm with priority on attribute values for distinct
attributes selected by calling the algorithm shown in Figure 4.b
In this way, the tree is first fully grown based on these conditions. Then under the pruning
process, we used two algorithms namely, NodeMerge and HeightBalance algorithms [11], which
help in enhancing the efficiency in case of dynamic pruning by avoiding the bushy growth of
decision trees, with HeightBalancePriority algorithm, gives priority and builds decision tree
considering all the classification rules, that C4.5 classifier had neglected.
2.4. Priority based Decision tree construction
The C 4.5 classifier gives preference to multi valued attributes, for which information entropy
is calculated and the attribute that has the highest information entropy is selected as root
node. This continues till the decision tree is fully grown.
Similarly in this priority based decision tree construction approach, Information
entropy is calculated for all attributes using the concepts explained in sections 2.1, 2.2, and
2.3. But attributes that are selected by the user is given higher preference, next the priority is
considered. During this, if the attribute selected by the user has less information entropy in
such a case the information entropy is to be replaced by attribute that has highest Information
entropy. This continues till all the attributes (which is of users interest) are selected and
replaced based on higher information entropy of other attributes which is of least important to
users.
Finally the decision tree is grown with all attributes that is of users interest, with all the
classification rules.
3. PROPOSED DECISION TREE ALGORITHMS
The Decision tree construction algorithm integrates attribute-oriented induction and
relevance analysis with modified version, of the C4.5 classifier is outlined in Algorithm
1. Algorithm 2, is used for merging of nodes as decision tree is built considering all the
nodes, is shown in Figure 3.
Algorithm 2. Merging of nodes at multiple levels of abstraction when decision tree is
constructed.
Input: All Nodes with attributes and attribute values given
Output: Merged decision tree is constructed.
NodeMerge( NodeData_A, NodeData_B)
{ Check priorities for node_A and node_B;
if both the priorities > checkpoint then
{ link_AB = remove_link_joining(NodeData_A, NodeData_B);
union = NodeData_A.merge_with(NodeData_B);

35
for (related_node: nodes_incident_to_either (NodeData_A, NodeData_B))
link_RA = link_joining (related_node, NodeData_A); link_RB = link_joining
(related_node, NodeData_B); disjoin (related_node, NodeData_A);
disjoin (related_node, NodeData_B);
join (related_node, union, merged_link);
}
else print (“ Node have high priority, cannot be merged”);
Call HeightBalance (union, new_link_AB); // calls Algorithm 3
}
Figure 3. Algorithm to merge nodes at multiple levels of abstraction during Decision tree
construction.
During the process of merging the nodes, the algorithm checks for assigned priorities for attribute
values of distinct attributes during decision tree construction. The primary disadvantage of
ordinary tree is that they attain very large heights in ordinary situation when large data sets are
given to classifier as seen in C4.5. To overcome this problem Heightbalance algorithm 3 is used
outlined in Figure 4 (a).
Decision tree constructed is prioritized, by giving preference to specific attribute considering
highest Information gain with the given algorithm 4, shown in Figure 4 (b).
Algorithm 3. Height balancing of decision tree.
HeightBalance (union, link_AB)
{ Check whether the tree is imbalanced or not;
if yes then
{ if balance_factor ( R ) is heavy
{ if tree’s right subtree is left heavy then
perform double left rotation;
else
perform single left rotation;
}
else if balance_factor( L ) is heavy
{ if tree’s left subtree is right heavy then
perform double right rotation;
else
perform single right rotation;
}
}
Print(“Tree is balanced”);
Check for path preservations;
}
Figure 4 (a). Height balancing for Decision tree
Algorithm 4. Height balancing for decision tree with priority.
Input: Nodes with Height balanced tree are given
Output: Prioritized Height balanced decision tree.

36
HeightBalancePriority (HeightBalance, Priority_Attribute)
{
Check for Priority in HeightBalance(Union,link_AB)
Evaluate (priority_attribute (Attributevalue1, .., Attributevaluen))
{
Switch(priority_attribute)
{
Case 1:
Priority_Attribute=’Attributevalue1’ //place it first
Check InformationGain //calculate Information Gain
Allocate the first Highest InformationGain //first priority Attrib.
Case 2:
Priority_Attribute=’Attribute value2’ //place it second
Allocate the second Highest Information Gain//second
priority attrib.
Case n:
Priority_Attribute=’Attributevaluen ’// place it n
Allocate the nth Highest InformationGain // nth priority
attrib.
Default:
Priority _Attribute = ‘Invalid’
}
}
Print(“Balanced Reconstructed Tree”)
Check for path preservations;
Generate Classification Rules;
}
Figure 4 (b). Prioritized Height balancing for Decision tree
Most of operations on decision trees are time consuming, which depends on height of the tree, so
it is desirable to keep the height as small as suggested in C4.5 leading to less classification rules,
where few attributes are missed out when the decision tree constructed. Some of the classification
rules found to be undesirable for the users, resulting in scalability and efficiency problem, which
will be discussed in Section 4 with an example.
4. PRIORITY BASED DECISION TREE CONSTRUCTION
Consider education dataset (edu_dataset) shown in Table 3, with following attributes
avg_edu_level, country, and income_level with attribute values of type {Illiterate, fouryrcollege,
twoyrcollege, graduateschool, elementaryschool, Phd}, {Cuba, USA, China, India} and {HIGH,
MEDIUM, LOW} respectively.

37
Table 3. Attributes with distinct attribute values.
Average Education
level
Country Income level
Illiterate Cuba Low
Fouryearscollege USA Medium
Fouryearscollege USA High
Fouryearscollege USA High
Twoyearscollege USA High
Graduate school China High
Elementary school Cuba Low
High school India Low
Fouryearscollege China Medium
Junior High China Low
Twoyearscollege India Medium
Fouryearscollege USA Medium
Ph. D India High
Illiterate Cuba Low
Twoyearscollege India Medium
When the decision tree is constructed using C4.5/J48 classifier [12] for the edu_dataset it has
pruned the attribute values of country attribute as shown in Figure 5. The Constructed decision
tree has attribute values of income_level only which reduces the scalability for large datasets.
The other important aspect is efficiency is not achieved, for proper evaluation of
classification rules; the resultant outputs obtained using J48 classifier of WEKA TOOL [12],
reveals that priority to be allocated for attribute values at each level of abstraction when the
decision tree is built.
Figure 5. Decision tree using C 4.5/ J48 classifier
This is achieved with our proposed Algorithms discussed, and for simplicity the working of
our algorithms shown, using modified DMQL query [9], with the example 4.1. The output
obtained is shown in Figure 6.
Example 4.1: Classification task.
Classify Decision_Tree
according to priority{country(India,USA,China,Cuba) attribute values}
in relevance to income_level
where attribute values for income_level count //leaf node(s)
income_level
USA(5/3) China(4/3)
=High
=High
India(3/1)
=medium

38
from edu_dataset
Figure 6. Decision tree with all attribute values obtained after DMQL query applied to income_level
attribute
It is observed, some of the nodes are pruned in C4.5 classifier when decision tree, is built as
shown in Figure 5. The outputs obtained after applying modified DMQL query that
constitutes our proposed algorithms, for decision tree construction for the same edu_dataset,
which is far better than earlier as shown in Figure 6, that considers all the attributes of users
interest.
5. PERFORMANCE EVALUATION
The classifiers like C4.5, ID3 and other classifiers (SLIQ, SPRINT) use decision tree
induction method which follows greedy top-down approach recursively. The decision
tree induction employs Attribute selection measure like Information Gain equations (1) and (2)
discussed in Section 2, that selects the test attribute at each node in the tree, during the pre-
pruning and post-pruning of branches in the decision tree construction. [1], [3], [4]. Outputs
obtained using WEKA tool, C4.5 classifier is shown in Figure 7.
Figure 7. Decision tree with multiple (IF…THEN) rules when the attribute avg_edu_level is
selected for education type
Attribute is represented by ellipse whereas actual attribute values are represented in rectangular
box with number specified throughout the figures drawn in this paper. The Decision tree
constructed in Figure 7, depicts that out of four distinct attributes i.e., avg_edu_level and
family_income_peryear attributes are only shown, remaining attributes like region and
income_level are pruned from the tree based on attribute selection measure of information gain. It
is also conveyed that the root node selected by C4.5 classifier, is avg_edu_level attribute always.
Family_income_per_year attribute can be calculated by counting the number of leaf nodes shown
in rectangular boxes, with two conditions only. The remaining conditions are skipped. Overall
the number of classification rules is reduced to nine in the decision tree.
Classification rules are retrieved by traversing the tree from root node to leaf nodes. To achieve
income_level
USA(5/3) China(5/3)
=High =High
India(4/2)
=medium
Cuba(2/2)
=low
Avg_edu_
level
Usa.east(3.2)
=fouryrscollege
Fam_inc_pyr
=twoyrscollege
=elementaryschool
Usa.south(2.1)
<=30400 >30400
Usa.middle(3.2) Cuba.north(1.0)
…..

39
this task of scalability and efficiency the NodeMerge and HeightBalance algorithms, applied by
considering the concepts of priority based decision tree construction, followed by AOI
(Attributed oriented induction), relevance analysis and multilevel mining technique as
discussed in Section 3, at each level considering all nodes. This is achieved with modified
DMQL query (constitutes our proposed algorithms), shown in example 5.1. The outputs
obtained are shown in Figure 8.
Example 5.1: Classification task.
Classify Decision_Tree
according to priority1{country(India,USA,China,Cuba)
attribute values}
according to priortiy2 {region(”India.east”, ”India.west”,
”India.south”,”USA.east”, ”USA.west”, ”USA.south”,
”USA.middle”, ”Cuba.north”, ”Cuba.south”,
”China.south”,”China.east”,”China.west”) attribute values}
in relevance to fam_inc_peryear
with attribute values for region count //leaf node(s)
from edu_dataset
All the Classification rules are retrieved by traversing the tree from root node to leaf nodes
accurately without missing a single rule, as seen in the earlier case shown in Figure 7. (previous
decision tree some of the attributes are skipped), In this way scalability for large databases with
multiple attributes at each level can be achieved efficiently with less classifier errors. The Minimum
Description Length (MDL) principle used to evaluate the cost of tree is discussed [13].
Figure 7. Priority based Decision tree based on above DMQL query
It is observed that in Figure 8, decision tree built, has considered multiple conditions for distinct
attributes and attributes values given, which is of users interest. The parameters that can be
compared with C4.5 classifier is shown in Table 4 with the proposed approach.
6. CONCLUSION
The methodology to improve C4.5 Classifier discussed, and the deficiencies exits in C4.5
classifier analyzed using WEKA tool by constructing decision trees with different types of data
sets, especially with large data sets, (for practical understanding, in our case we have taken
T4I5K3N8 education data set) resulted in few classification rules as shown in Section 4,
using C4.5 classifier.
Avg_edu_
level
Usa.east(3.1)
=fouryrscollege
Fam_inc_py
r
=twoyrscollege
=elementaryschool
Usa.south(2.1)
<=30400 >30400
Usa.middle(3.1) Cuba.north(1.1)
…..
Fam_inc_py
r
<=30400
India.south(3.1)
India.west(3.1)
Fam_inc_py
r
Usa.middle(2.1)
China.east(3.1)
USA.west(3.1)
<=30400

40
To achieve scalability and efficiency in decision trees, many Classification rules are retrieved.
NodeMerge and HeightBalance algorithms, applied by considering the concepts of priority based
decision tree construction i.e HeightBalancePriority algorithm, as discussed in Section 3, at
each level that consider all nodes. Modified DMQL query, is used for better understanding as
discussed in Section 4 and 5 with many examples by taking education data set (edu_data set).
Performance is evaluated based on parameters mentioned in Table 4. The other parameters like
classification of correctly classified instances and incorrectly classified instances can also be
evaluated using precision and recall measures.
Table 4. Comparison of C 4.5 /J48 classifier with the proposed approach
(HeightBalancePriority Algorithm).
Our future approach is to use the constructed priority based Decision tree, for Predicting the
missing and unknown values in very large data sets, and evaluate them using Confusion matrix.
REFERENCES
[1] J. R. Quinlan, 1996. Improved use of continuous attributes in C4.5 for Journal of Artificial
Intelligence Research, 4:77-90.
[2] Xindong Wu, Vipin Kumar,J.Ross Quinlan, Joydeep Ghosh, Qiang Yang, Hiroshi
Motoda,Geoffrey, J. McLachlan, Angus Ng, Bing Liu, Philip S. Yu, Zhi-Hua Zhou,Michael
steinbach, David J. Hand, and Dan Steinberg, 2007. A survey paper on Top 10 algorithm in data
mining, for Springer.
[3] M. Mehta, R. Agrawal, and J. Rissanen, 1996. SLIQ: A fast scalable classifier for data mining. In
Proceeding of International Conference on Extending Database Technology (EDBT’96), Avignon,
France.
[4] H J. Shafer, R. Agrawal, and M. Mehta, 1996. SPRINT: A scalable parallel classifier for data mining.
In Proceedings of the 22nd International Conference Very Large Databases (VLDB), pages 544-555,
Mumbai, India.New York: Springer-Verlag, 1985, ch. 4
[5] J. Han, and M. Kamber, 2006. Data Mining Concepts and Techniques, Elsevier Publishers.
[6] D. H. Freeman, Jr., 1987. Applied Categorical Data Analysis, Marcel Dekker, Inc., New York.
[7] Micheline Kamber, Lara Wistone, Wan Gong, Shan Chen, and Han, 1997. Generalization and
Decision Tree Induction: Efficient classification in Data Mining, for IEEE.
Parameter
Classifier
Measure used for
pruning and decision
tree construction
Number of
attributes used
for decision
tree
construction
with priority
Multiple conditions Selection
of attribute
with
attribute
values
Pruning
and
DMQL support
C4.5/J48 Information Entropy Few nodes
without priority
Few Classification
rules
Not
supported
Top-down
approach
recursively, Pre
and post pruning,
no DMQL query
support
Proposed
approach
with Priority
Highest Info. gain &
Uncertainty co-
efficient, with
threshold coefficient
and priority
All nodes
considered with
priority using
HeightBalance
Priority
algorithm
Many
Classification rules,
helps to satisfy user
specified QUERY
supported NodeMerge and
Height balance
algorithms with
Dynamic pruning,
matching user
specified DMQL
query

41
[8] Tomoki Watanuma, Tomonobu Ozaki, and Takenao Ohkawa,2006.Decision Tree Construction
from Multidimensional Structured Data, Sixth IEEE International Conference on Data Mining
(ICDMW'06) – Workshops, Hong Kong, China.
[9] J. Han, Y. Cai, and N. Cercone, 1993. Data driven discovery of quantitative rules in relational
databases, for IEEE Transactions Knowledge and data Engineering,5:29-40.
[10] L. B. Holder 1995. Intermediate decision trees,” In Proc. 14th Int., Joint Conf. on Artificial
Intelligence, pages 1056–1062, Montreal, Canada.
[11] Mohd Mahmood Ali and Lakshmi Rajamani, 2009. Decision Tree Induction: Data Classification
using Height-Balanced Tree, In Proceedings of International Conference of Information and
Knowledge Engineering at(IKE ’09), U.S.A.
[12] Mark Hall, Eibe Frank, Geoffrey Holmes, Bernhard Pfahringer, Peter Reutemann, and Ian H.
Witten, 2009. The WEKA Data Mining Software: An Update, for SIGKDD Explorations, Volume 11,
Issue 1.
[13] J.R. Quinlan and R.L. Rivest, 1989. Inferring Decision Trees Using the Minimum Description Length
Principle, Journal of Information and Computation, ACM volume 80, pp. 227-248, 1989.
[14] Mohammed Mahmood Ali, and Lakshmi Rajamani, 2012. Decision Tree Induction: Priority
Classification, for IEEE. In Proceedings of International Conference Advances in Engineering,
Science and Management (ICAESM ’12), India.
Authors
Mohd. Mahmood Ali, working as an Assistant Professor in the Dept. of C.S.E.,
Muffakhamjah College of Engg. & Tech., Hyderabad, India is a research scholar pursuing
part-time Ph.D. from Osmania University, Hyderabad, A.P. India under the guidance of
Dr. Lakshmi Rajamani, Prof CSE, College of Engg., Osmania University, Hyderabad
India.mahmoodedu@gmail.com
Mohd. S. Qaseem, working as Associate Professor in the Dept. of C.S.E / IT, Nizam Institute of Engg. &
Tech., Hyderabad, India is a research scholar pursuing part-time Ph.D. from Acharya Nagarjuna University,
Guntur, A.P. India under the guidance of Dr. A. Govardhan, Prof CSE , DE JNTU Hyderabad India.
ms_qaseem@yahoo.com
Dr. Lakshmi Rajamani, Professor, Dept. of C.S.E., Univ. College of Engg., O.U., Hyderabad, India. She
has published more than 40 papers in reputed journals and international conferences includes IEEE, ACM,
Springer, Inder science and others. She has supervised more than 7 Ph.D Scholars. She is also an active
member of IEEE and ACM.drlakshmiraja@gmail.com
Dr. A. Govardhan, Professor and Director of Evaluation, JNTU Hyderabad, India. dejntuh@jntuh.ac.in

EXTRACTING USEFUL RULES THROUGH IMPROVED DECISION TREE INDUCTION USING INFORMATION ENTROPY

More Related Content

Similar to EXTRACTING USEFUL RULES THROUGH IMPROVED DECISION TREE INDUCTION USING INFORMATION ENTROPY (20)

More from ijistjournal (20)

Recently uploaded (20)

EXTRACTING USEFUL RULES THROUGH IMPROVED DECISION TREE INDUCTION USING INFORMATION ENTROPY