A Comprehensive Analysis on Associative Classification in Medical

Document technical information

Format pdf
Size 257.6 kB
First found May 22, 2018

Document content analysis

Category Also themed
Language
English
Type
not defined
Concepts
no text concepts found

Persons

Organizations

Places

Transcript

Indian Journal of Science and Technology, Vol 8(33), DOI: 10.17485/ijst/2015/v8i33/80081, December 2015
ISSN (Print) : 0974-6846
ISSN (Online) : 0974-5645
A Comprehensive Analysis on Associative
Classification in Medical Datasets
D. Sasirekha1* and A. Punitha2
Department of Computer Science, Bharathiyar University, Coimbatore - 641046, Tamil Nadu, India;
[email protected]
2
Department of Computer Applications, Queen Mary’s College, Chennai - 600004, Tamil Nadu, India;
[email protected]
1
Abstract
Association rule mining along with classification technique is capable of finding informative patterns from large data
sets. Output of this technique is of the form if-then which is easy to understand for the end users and also for prediction.
Termed as Associative Classification, it is having wide application on medical domain in diagnosing diseases and to
analyze medical datasets which are unstructured, heterogeneous, incomprehensible and voluminous. Analysing these
data allows physicians to predict the diseases as well as to take vital decisions. This paper presents a detailed study
on associative classification and the phases of associative classification procedure. Several associative classification
methods viz., CPAR, MMAC, CAR, CBA, CMAR, etc. along with their merits and demerits are also presented in a lucid
manner.
Keywords: Association, Associative Classification, Classification, Knowledge Discovery, Medical Datasets
1. Introduction
Medical data contains huge volume of information in an
unstructured format. Data mining discovers insightful,
interesting and novel patterns which are descriptive,
understandable and predicative from large amount
of data1. Data mining includes core techniques such
as association, classification, clustering, prediction,
combination, etc., that are used for different type of mining
and data recovery operations. Association is a technique
used to know the correlation between two or more items.
Association rule mining defines all the rules that exist
in the data set that satisfies some amount of support
and confidence. The best application of association rule
mining is market basket analysis which investigates the
purchase behaviour of customers2. Association rule
mining tries to find the group of items that are frequently
purchased in the presence of other items in the shopping
cart. This group of items is said to be frequent itemset.
On the other hand, classification is also one of the most
* Author for correspondence
important techniques in data mining. Constructing fast
and accurate classifier for large data is an important task
in knowledge discovery. There are several classification
techniques such as Naive Bayesian classifier, C4.5, PART,
Prism, IREP, etc. Normally, classification techniques
produce small subsets of rules and there is a great chance
that detailed rules that are vital in decision making may
get skipped3.
In association rule mining, the target of knowledge
discovery is not predetermined, whereas classification
rule mining works by fixing one and only one
predetermined target4. In order to make it convenient
and to save time in knowledge discovery process, these
two techniques can be integrated. A new approach that
integrates association rule mining and classification is
said to be Association Classification5. This fusion is done
by considering with minimum subset of association rules
called Class Association Rules (CARs). This new approach
yields higher accuracy5. There exist many associative
classification techniques.
A Comprehensive Analysis on Associative Classification in Medical Datasets
X. Yin and Jiawei Han6 presented Classification
based on Predictive Association Rules (CPAR), which
combines the advantages of both associative classification
and rule based classification. CPAR adopts greedy
method to generate rules from large training data. CPAR
generates and tests more rules than ordinary rule-based
classification. Wenmin Li et al.7 proposed accurate and
efficient classification based on multiple class association
rules. The method extends Frequent Pattern (FP) growth
mining method, constructs FP-tree and mines large
databases in an efficient way. CPAR is consistent and
highly effective.
Experimental results carried out by many studies show
that classification based association rule mining is highly
effective and constructs strong predictive and accurate
classifiers than traditional classification methods6,7.
2. Association Rule Mining
Association Rule Mining (ARM) is a tool used to find
interesting associations and correlations among large set
of data items which is a strong mechanism used in market
basket analysis. Association rules shows attribute value
conditions that occur frequently together in a given data
set. Association rule mining finds group of items that are
frequently sold together along with the presence of other
items in customers shopping cart. Let D be a database
which stores transaction such that D = {t1, t2, t3…. tn} and
I be the set of items. Set of items, called as itemset and an
itemset with k-items is said to be k-itemset. The support
of an itemset is the number of transactions in D where
itemset occurs as subset and it is denoted by σ. Itemset
is said to be frequent or large if support value is greater
than user specified minimum support8. Association rule
is of the form X→Y, where X and Y are disjoint itemsets.
The support for the rule X→Y, is the probability of
transactions containing both X and Y. The confidence of
the association rule X→Y is the conditional probability
that atransaction contain Y, given that it contains X. An
association rule is frequent and strong if its support is
greater than minimum support and its confidence is
greater than minimum confidence respectively,
s(X  Y)
N
s(X  Y)
confidence(X ® Y)=
s(X)
support(X ® Y)=
The ARM process works in two phases, with finding
2
Vol 8 (33) | December 2015 | www.indjst.org
all frequent itemsets having minimum support and
confidence in the first phase, and generating strong
association rules from the frequent itemsets in the second
phase. Apriori is the most classic association rule mining
algorithm which achieves good level of performance.
However, ARM is having certain drawbacks such as
• Holding large number of candidate itemsets which is
costly.
• Scanning the database multiple times to check number of candidate itemsets, which is again costly during
runtime.
• Items with high confidence and low support may get
ignored.
• Setting appropriate parameter for mining algorithm.
• ARM discovers too many rules.
• Discovery of poorly understandable rules.
3. Classification
It is the process of predicting a class label for a given
unlabeled point. It involves examining the features of new
objects and trying to assign one predefined set of classes.
Building a model from classified objects in order to
classify previously unseen objects as accurately as possible
is the goal of classification9. Devising a procedure for
classification in which exact classes are known in advance
is termed as pattern recognition or supervised learning.
There are many cases of unsupervised learning in which
classification class is not known in advance. Three main
standards of classification techniques include, statistical
based classification, machine learning based classification
and neural network based classification16.
All the above classification approaches work by
adopting divide-and-conquer, separate-and-conquer and
by statistical approaches. Various classification algorithms
have been developed such as PART, RIPPER, Prism, etc.
A traditional classification problem can be defined as
follows: Let D denotes the domain of all possible training
instances, let Y denotes list of class labels and H denotes
set of classifiers. The objective is to find a classifier h such
that h ∈ H which minimizes the probability that
h(d) = y for each test case (d, y)
Apart from various advantages, classification
techniques contain certain drawbacks also. Few
classification algorithms generate empty branches or
insignificant branches10. Few algorithms require long
searching time and more memory. Few classification
algorithms may not perform well with very few records
Indian Journal of Science and Technology
D. Sasirekha and A. Punitha
as they need large number of data to achieve good results.
Moreover, majority of the classification algorithms are
costly.
4. Associative Classification
Associative classification is a type of classification approach
that adopts association rule mining and classification
rule mining to build classification models11. It is a special
kind of ARM where target attribute is considered in
rule’s right hand side. In associative classification, the
entire data set is divided into two parts with 70% of data
used for training and 30% for testing the accuracy of the
classifier11. Associative classification contains various
advantages such as
• Training the data is very efficient regardless of the size
of training set.
• High dimensionality training sets can be handled easily and assumptions on dependent and independent
attributes are not needed.
• Classification is done in quick manner.
• The generated rules are easy to understand by humans.
Associative classification works in three steps,
• Generate a set of association rules from training set
with certain support and confidence threshold as candidate rules.
• Prune the discovered rules that may introduce over
fitting.
• Perform classification to make prediction for test data
and measure the accuracy of the classifier.
Associative classification uses mainly minimum
support and minimum confidence13. The data in associative
classification can be represented in three ways such as
horizontal, vertical and set theory data representation.
Horizontal representation is like a format adopted in
ARM. In this format, a database consists of transactions,
with each transaction being identified with transaction
identifier and list of items in the transaction12. The data
is represented in rows, with each row representing a
transaction. Classification Based Association (CBA) uses
this type of representation.
Vertical data representation transforms training data
into a table form which contains transaction identifier
and simple intersections are employed to discover
frequent itemsets to produce CARs12. Associative
classification algorithms such as Multi-Class Multi-
Vol 8 (33) | December 2015 | www.indjst.org
Label Associative Classification (MMAC) and MultiClass Classification based Association Rules (MCAR)
use vertical data representation. During rule discovery,
frequent items of N-1 size discovers the possible frequent
items of size N. Rough set theory data representation is
based on discarding redundant attributes from training
data sets and selects reducts of attribute-value pairs that
can represent the complete training data set in a decision
table context14. ROSETTA selects reducts of attributevalue pairs. Reduct is a subset of attribute values and class
attribute that represents the whole table.
4.1 Associative Classification Methods
4.1.1 CBA
Classification Based on Association Rules (CBA) selects a
sub set of rules to form the classifier model. CBA works
in two phases viz., rule generation and classifier building.
The perfect classifier is built using Class Association
Rules (CARs). CBA is a simple algorithm which prefers
database to be inside the main memory. There are
several associative classification algorithms that use CBA
properties to find frequent itemsets and to generate class
association rules.
Zhitong Su24 adopted CBA for discovering informative
association rules. To select and rank small subset of
high quality rules, entropy based associative classifier
is built. This classifies uses information gain parameter
to select and rank the rules. Shekhawat and Dhande25
adopted CBA methodology and Neural Network based
Associative Classification (NNAC) system is presented.
NNAC improves the efficiency of the classifier.
4.1.2 CMAR
CMAR stands for classification based on multiple
association rules. CMAR performs weight analysis using
multiple strong association rules. The weight analysis is
performed with Chi-Square (χ2) method. This determines
the strength of the association rule under both support
and class distribution28. CMAR also works in two phases
with rule generation in first phase and class distribution
in second phase.
In order to improve efficiency and accuracy,
Classification Rule Tree (CR-Tree) data structure is
used. The new data structure is an extension of Frequent
Indian Journal of Science and Technology
3
A Comprehensive Analysis on Associative Classification in Medical Datasets
Pattern Growth (FP-tree), and it compactly stores and
retrieves large number of rules for classification. In order
to speed up the classification process, CMAR uses variant
of FP-growth algorithm, which is faster than apriori-like
methods28.
4.1.3 CARGBA
It stands for classification based on association rule
generated in a bidirectional approach. CARGBA works
in two phases. In the first step, it generates a set of high
confidence rules of smaller length27. Then this set is
increased by adding high confidence rules of higher
length. In the second phase, specific rules are generated
without any support and pruning. This result is a better
mixture of class association rules. Since all the rules are
not used in classification, the second phase of CARGBA
builds a classifier called CARGBA classifier.
4.1.4 CPAR
CPAR stands for classification based on predictive
association rules. CPAR uses expected accuracy measure
to evaluate each rule to avoid over fitting problem. CPAR
follows the basic idea of First Order Inductive Learner
(FOIL) algorithm in rule generation29. It seeks for the
best rule condition that can bring maximum gain among
available dataset. Once the condition is identified, weights
of positive examples associated with it will be deteriorated
by a multiplying factor.
During classifier building process, a common error
occurs, called misclassification penalty. This error can be
minimized by using modified CPAR (M-CPAR)26. CPAR
combines the advantages of associative classification and
traditional rule based classification. CPAR generates rules
directly from training data by using greedy algorithm.
4.1.5 CARPT
Classification Algorithm based on association rules. This
method is more preferable if frequent rules need to be
generated directly with generating frequent itemsets. Here,
the data of any form is converted into two-dimensional
array. The horizontal position of two dimensional array
represents item number and property types. The vertical
position represents transaction number30.
CARPT uses Trie-tree to build accurate classifier. Trietree influences two properties for generating frequent
rules. First property states that “if a sub tree takes a non-
4
Vol 8 (33) | December 2015 | www.indjst.org
frequent bucket for root nodes, then all the buckets of the
sub tree are not frequent”. Second property states that “if
<i1, i2, …, in> is the frequent item set. Then there cannot
be frequent item set which contains two or more items
taken for a prefix”. Usage of trie-tree reduces the cost of
the query time thereby improving efficiency30.
4.1.6 HMAC
Hierarchical Multi-label Associative Classification
(HMAC) is another classification technique performed
on hierarchically arranged data. Sawinee Sangsuriyun et
al.31 presented a paper on HMAC using negative rules.
Here, the Rules of Positive Set (RPS) are generated first.
The generated rules are pruned so that rules without
minimum support and confidence are removed.
Pruning is done using two methods such as Pearson’s
correlation coefficient pruning and redundant rules
pruning. For a given set of coefficient, φ is used to select
rules with user specified threshold. HMAC classifier
is built, which selects a sub set of high quality rules.
HMAC classifier also predicts the class label for the object
by analyzing the subset rules31. The efficiency and set
similarity of this algorithm is tested with F-measure and
Jaccard’s coefficient.
4.1.7 ACCF
ACCF stands for Associative Classification based on
Closed Frequent itemsets. It was proposed by Li et al.32
and it works in two phases viz., rule generation and
classifier construction. ACCF is the extended version
of CHARM algorithm. ACCF discovers a set of Closed
Frequent Itemsets (CFIs) along with their tid (transaction
identifier) sets and class labels.
The support and confidence for each rule is calculated
by using tidsets for both items and corresponding class
labels. A set of CARs is generated by evaluating the
rules and selecting the rules with minimum support and
minimum confidence. The generated CARs are ranked
according to their support, confidence and rule length.
ACCF starts with first ranked rule and applies to tidset
for classification32.
4.1.8 GARC
Gain based Association Rule mining (GARC) uses
extended association rule mining to construct a classifier33.
It works in three steps. First it discovers candidate itemsets
Indian Journal of Science and Technology
D. Sasirekha and A. Punitha
by applying information gain measure. The candidate
itemsets with best-split attribute value are generated. By
this way, redundant and unwanted candidate itemsets are
removed. In the second step, it combines rules generation
and frequent itemset generation process. In the third step,
it produces compact set of rules that are short and easy to
understand.
The classifier built using GARC can be applied to
discrete and continuous data. The performance of GARC
can be enhanced by applying pruning strategies33. GARC
classifier produces better results in terms of efficiency,
accuracy and understandability.
4.1.9 ACN
Associative Classifier with Negative rules (ACN) is a
different associative classification method that mines
relatively large set of negative association rules first. Then
positive and negative association rules are used to build
a classifier. The advantage of generating negative rules is
that, good number of negative rules can be replaced with
weak positive rules. Hence, there will be a reduction in
number of inaccurate positive results in the final classifier
thereby increasing the accuracy34.
Generally computational cost of mining negative
association rules is high and it will become still higher
if positive association rules are also need to be mined.
ACN uses the variant of Apriori algorithm to address the
cost issue and tries to generate both forms of rules with
low overhead. ACN works in two phases. First it designs
an ACN rule generator and then it builds classifier. ACN
uses several pruning strategies to cut down the number of
rules generated34. ACN is highly consistent, accurate and
effective on various kinds of databases.
4.1.10 Comparison of Various Associative
Classification Methods
Table 1. Comparison of associative classification methods
Method
CMAR
•
•
CPAR
•
By pruning process, CMAR selects only high quality rules for classification.
•
CMAR is superior to C4.5 and CBA interms of
accuracy and it is also scalable.
Redundancy is avoided since each is generated by
comparing with the set of already generated rules.
•
•
CBA
CARGBA
Advantages
It finds frequent patterns and generates association •
rules in one step.
•
Since CR-tree data structure is used, both accuracy
and efficiency is improved.
•
It generates high quality predictive rules directly
from dataset.
Simple algorithm, finds valuable rules.
CPAR is more complex to understand as well as to
implement.
•
Usage of greedy algorithm to train the dataset adds additional computational overhead to the algorithm.
Training the dataset often generates huge set of rules
leading to redundancy.
•
Capable of handling data in table form as well as in
•
transaction form.
•
It does not require the whole dataset to be fetched
into main memory.
Capable of handling various kinds of databases,
and the method is highly consistent and effective.
•
Rules are produced without any support pruning.
Vol 8 (33) | December 2015 | www.indjst.org
Performing weighted analysis adds substantial computational load to the algorithm.
•
•
•
Disadvantages
Since it searches for only high quality rules, it is slower.
Classification is based on parameter called confidence,
which can be biased sometimes.
•
Multiple scans are needed to generate rules which makes
the algorithm slow.
•
Overall accuracy of the algorithms is less.
Indian Journal of Science and Technology
5
A Comprehensive Analysis on Associative Classification in Medical Datasets
CARPT
•
Storing data in vertical data format reduces number of scans to the database.
•
The data should be in two dimensional array of vertical
data format.
GARC
•
•
Time and space is saved effectively.
It gets rid of many candidate itemsets in the training phase itself.
•
•
It removes frequent items.
Only few rules are generated.
ACCF
•
•
Handles both continuous and discrete datasets.
No redundant rules are generated.
•
Taking additional time for learning the frequent itemset.
•
Usage of coverage pruning algorithm discards rules
that does not cover atleast one training case.
It is time efficient and achieves better accuracy.
•
ACN
•
•
HMAC
•
•
Too many negative rules are generated with low
cost and these negative rules control the inclusion
of inaccurate positive rules.
Improved accuracy is achieved since negative
redundant rule pruning is used.
Capable of handling complex data.
4.2 Rule Ranking Procedures
The generated class association rules are well-organized
using support and confidence, and rules with high
support and confidence is ranked on top. There are several
methods available to rank the generated class association
rules. CBA algorithm uses support, confidence and
antecedent length to arrange the discovered rules15. Rules
are arranged based on confidence. If two or more rules
are having same confidence, then they are arranged by
their support. If the support values are also same, then
their antecedent lengths are used for arranging. If the
antecedent lengths are also identical, then randomly
rules are arranged15. Other algorithms such as CARGBA,
Associative Classification based on Closed Frequent
itemsets (ACCF) are also using this approach for
arranging rules. It is also possible to order the generated
rules without the above three parameters. Instead, a new
parameter is used which is called as class distribution,
which reflects the number of items in a class, is used to
arrange the generated rules.
Baralis and Torino11 proposed lazy associative
classification algorithms which uses antecedent lengths,
which is the opposite of CBA rule ranking practise. Here
the rules which hold more attributes are ranked on top
and these rules are called as specific rules. Information
gain, another parameter used to rank the generated
rules. Information gain is a mathematical measure
that represents how well a given attribute is used for
6
Vol 8 (33) | December 2015 | www.indjst.org
•
It calculates support and confidence for negative rules
which is unnecessary and adds computation overhead.
Rule generation is based on combination of two scoring
methods viz., F-measure and Jaccard’s measure which
leads to additional computational overhead.
classification. To calculate information gain of an
attribute, it is important to calculate entropy.
D=å -Pk log 2 Pk
Where Pk is the probability of class k belongs to D.
The information gain of an attribute is calculated by
G = Entropy (D ) - å ((| Da | / | D |) * Entropy (Da ))
Where Da is the subset of an attribute which has the
value a, modulus of Da is the number of data cases in Da
and modulus of D is the number of data cases in D.
4.3 Building the Classifier and Pruning
Classifier is the set of rules that are built from training set.
Since a classifier builds many rules, result is delayed in
classification process. Therefore it is necessary to remove
useless and redundant classifier16. And also this removal
will increase the speed of the classification process.
Database coverage is a pruning method which checks
for the rules covering at least an object. If so, it is added
in the classifier and its corresponding training object
is deleted from the training dataset. This procedure is
repeated until all training objects are deleted or all the
rules are examined. If the training data set is not empty at
the end of pruning, all the remaining cases are put under
default class rule. This procedure is adopted by several
associative classification algorithms such as CBA, CMAR,
CAAR, ACN and ACCF.
Indian Journal of Science and Technology
D. Sasirekha and A. Punitha
High precedence method iterates over the ranked rules
starting with highest rank. All the training cases covered
are discarded and the rule is inserted into the classifier.
Rule that does not cover training case are removed. High
classification pruning method selects a rule that partially
covers a training case and has a common class to that of
training case17. Based on this, a training case is inserted
or deleted from the classifier. The whole procedure is
repeated until training data becomes empty.
Pruning can also be done using mathematical
calculations. The decision to replace a sub-tree with a leaf
node or to keep as such is done by calculating pessimistic
error estimation measure over training data set. CBA uses
this approach. Chi-Square test can also be used to decide
a rule relevancy for its inclusion in class association rules
set.
CMAR uses this approach for pruning the rules
set. Redundant rules are removed by using long rules
pruning which eliminates long rules with confidence
values larger than their subset rules18. The rules that may
lead to misclassification on the training data set can be
removed by using lazy method. In this method, every
rule that covers a training data case and have both the
same class, the rule is inserted in the primary rule set else
deleted. Conflicting rules, which may cause the classifier
to classify the test case, must be removed.
5. A
ssociative Classification in
Medical Datasets
Medical data mining is an emerging technology in medical
field that solves the traditional medical problems such as
congestion, long wait time and delayed patient case. This
technique helps physicians to make accurate diagnosis of
diseases. Generally medical data set are widely distributed,
in unstructured format, heterogeneous in nature and
in huge volume. These data need to be organized in a
form which is understandable. Advantage of using data
mining in medical domain is to improve the accuracy
of the output with huge amount of data. Medical data
mining has great potential for exploring hidden patterns
among medical data sets. Among various data mining
techniques, associative classification technique is having
a very good application on medical domain. Most of the
attributes in medical data sets are often associated with
quantitative domains such as Body Mass Index (BMI),
age, Blood Pressure (BP), etc.
Vol 8 (33) | December 2015 | www.indjst.org
Rafel Rak et al.19 proposed multi-label associative
classification on medical documents from Medical
Literature Analysis and Retrieval System Online
(MEDLINE). This method performs classification on
medical repository MEDLINE. It is based on associative
classification which considers multi label features. Akhil
Jabbar et al.20 proposed heart disease prediction system
using associative classification and genetic algorithm.
This method performs high level rules that are accurate
and comprehensible, and contains high interesting value.
K. Ruth Ramya et al.21 proposed class based approach
for medical classification of chest pain. The class label is
used in classification to minimize the searching space.
This method also synchronizes the rule generation and
classifier building phase.
Genetic algorithms are adaptive procedures that
maintains a population of potential solutions to candidate
set problem. The genetic algorithm contains three
operators such as selection, crossover and mutation.
These operators allow algorithm to explore search space.
Genetic algorithms can be implemented in two ways:
through external support and direct application over
analysis. The reason for preferring genetic algorithm in
data mining is that more attributes and more observations
can be handled. However, genetic algorithm can be
implemented only when the data is in discrete structures.
Fitness function used by selection operator in genetic
algorithm provides support notion for an association.
It defines figure of merit divided by utility measure of a
sample in numerical format. The fitness function works
in two phases; with finding associations of small support
in the first phase and finding associations of large support
in the second phase. The fitness function can be expressed
using the below formula
T
T
æ
æ
-10xSF ö÷
+ 10xSF ö÷
F = çç(1 - S ) x 10
+ 2x ççS x 10
÷
÷
÷
èç
ø
èç
ø÷
T
T
Where S represents the support, T represents total
number of features and SF represents number of selected
significant features. Fitness function is a flexible way of
expressing model criteria and tradeoffs among multiple
objectives.
Genetic algorithms is having wide applications
in medical domain. Asha Gowda Karegowda et al.21
presented a paper on applications of genetic algorithm
for medical diagnosis of PIMA Indian diabetes. This
model integrates genetic algorithm and back propagation
Indian Journal of Science and Technology
7
A Comprehensive Analysis on Associative Classification in Medical Datasets
network to diagnose diabetes. Erhan Elveren and Nejat
Yumusak22 proposed tuberculosis diagnosis system
using genetic algorithm. The method works with neural
network trained data and genetic algorithm. Prem Pal
Singh Tomar and Ranjit Singh23 presented evolutionary
continuous genetic algorithm for clinical decision support
systems. This method, called Medical Multimedia based
Clinical Decision Support System (MM-CDSS), supports
diagnosis for four major heart diseases using patient’s
symptoms.
6. Conclusion
This paper presents ananalysis on associative classification
and various traditional methods of associative
classification. Performed review outlines the merits
and demerits of traditional associative classification
algorithms. For example, CARPT method saves time
and space, whereas it removes frequent items during
classification process. It is concluded from the study
that if the objective of the classification is to generate
high quality rules without redundancy, CPAR and
CBA can be used. If the objective is to generate high
quality accurate rules, CMAR is the best option. For
performing classification on heterogeneous data sets,
it is better to use CARGBA. Associative classification
can also be performed by generating negative rules.
ACN and ACCF methods use this concept and they are
accurate in generating association rules. The variations
of CBA algorithm such as HMAC and GARC, performs
associative classification with higher accuracy. This paper
reviews about associative classification on medical data
sets. It is obvious from the study that inclusion of genetic
algorithms while performing associative classification on
medical data sets will produce better results.
7. References
1. Kononenko, Igor, Kukar M. Machine learning and data
mining: Introduction to principles and algorithms. Horwood Publishing; 2007.
2. Chen MC, Chiu AL, Chang HH. Mining changes in customer behavior in retail marketing. Expert Systems with
Applications. 2005; 28(4):773–81.
3. Thabtah F, Cowling P, Peng Y. MCAR: Multi-Class Classification based on Association Rule. The 3rd ACS/IEEE
International Conference on Computer Systems and Applications; 2005.
8
Vol 8 (33) | December 2015 | www.indjst.org
4. Rygielski C, Wang JC, Yen DC. Data mining techniques for
customer relationship management. Technology in Society.
2002 Nov; 24(4):483–502.
5. Weiyang L, Alvarez SA, Ruiz C. Efficient adaptive-support
association rule mining for recommender systems. Data
Mining and Knowledge Discovery. 2002; 6(1):83–105.
6. Yin X, Han J. CPAR: Classification based on Predictive Association Rules. SDM. 2003; 3:331–5.
7. Li W, Han J, Pei J. CMAR: Accurate and efficient classification based on multiple class-association rules. Proceedings
IEEE International Conference on Data Mining-ICDM;
San Jose, CA. 2001. p. 369–76.
8. Tao F, Murtagh F, Farid M. Weighted association rule mining using weighted support and significance framework.
Proceedings of the Ninth ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining,
ACM; 2003. p. 661–6.
9. Kirkos E, Spathis C, Manolopoulos Y. Data mining techniques for the detection of fraudulent financial statements.
Expert Systems with Applications. 2007; 32(4):995–1003.
10. Zheng Y, Liu P, Lei L, Yin J. R-C4. 5 decision tree model and
its applications to health care dataset. 2005 Proceedings of
International Conference on Services Systems and Services
Management (ICSSSM’05); 2005. p. 1099–103.
11. Baralis E, Chiusano S, Garza P. A lazy approach to associative classification. IEEE Transactions on Knowledge and
Data Engineering. 2008; 20(2):156–71.
12. Yu G, Li K, Shao S. Mining high utility itemsets in large
high dimensional data. Proceedings of the 1st International Conference on Forensic Applications and Techniques in
Telecommunications, Information, and Multimedia and
Workshop. ICST (Institute for Computer Sciences, Social-Informatics and Telecommunications Engineering);
2008. p. 47.
13. Lin W, Alvarez SA, Ruiz C. Efficient adaptive-support association rule mining for recommender systems. Data Mining and Knowledge Discovery. 2008; 6(1):83–105.
14. Thabtah F. A review of associative classification mining.
The Knowledge Engineering Review. 2007; 22(01):37–65.
15. Zaki MJ. Mining non-redundant association rules. DataMining and Knowledge Discovery. 2004; 9(3):223–48.
16. Kotsiantis SB, Zaharakis I, Pintelas P. Supervised machine
learning: A review of classification techniques. Emerging
Artificial Intelligence Applications in Computer Engineering; 2007. p. 3–24.
17. Nauck D, Kruse R. Obtaining interpretable fuzzy classification rules from medical data. Artificial Intelligence in Medicine. 1999; 16(2):149–69.
18. Scheffer T. Finding association rules that trade support optimally against confidence. Principles of Data Mining and
Knowledge Discovery. Heidelberg: Springer Berlin; 2001. p.
424–35.
19. Rak R, Kurgan L, Reformat M. Multi-label associative
classification of medical documents from medline. IEEE
Proceedings Fourth International Conference on Machine
Learning and Applications; 2005. p. 177–86.
Indian Journal of Science and Technology
D. Sasirekha and A. Punitha
20. Jabbar MA, Deekshatulu BL, Chandra P. Heart disease prediction system using associative classification and genetic
algorithm. International Conference on Emerging Trends
in Electrical, Electronics and Communication Technologies; 2013. p. 183–92.
21. Karegowda AG, Manjunath AS, Jayaram MA. Application
of genetic algorithm optimized neural network connection
weights for medical diagnosis of pima Indians. International Journal on Soft Computing. 2001; 2(2):17–21.
22. Elveren E, Yumusak N. Tuberculosis disease diagnosis using artificial neural network trained with genetic algorithm.
Journal of Medical Systems. 2011; 35(3):329–32.
23. Tomar PPS, Singh R. Evolutionary continuous genetic algorithm for clinical decision support system. African Journal
of Computing and ICT. 2013; 6(1):127–46.
24. Su Z, Song W, Cao D, Li J. Discovering informative association rules for associative classification. IEEE International Symposium on Knowledge Acquisition and Modeling
Workshop; Wuhan. 2008. p. 1060–3.
25. Shekhawat PB, Dhande SS. A classification technique using
associative classification. International Journal of Computer Application. 2011; 20(5):20–8.
26. Ganganwar V. An overview of classification algorithms for
imbalanced datasets. International Journal of Emerging
Technology and Advanced Engineering. 2012; 2(4):42–7.
27. Jiang Y, Shang J, Liu Y. Maximizing customer satisfaction
through an online recommendation system: A novel as-
Vol 8 (33) | December 2015 | www.indjst.org
sociative classification model. Decision Support Systems.
2010; 48(3):470–9.
28. Ye Y, Li T, Jiang Q, Wang Y. CIMDS: Adapting postprocessing techniques of associative classification for malware detection. IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews. 2010; 40(3):298–307.
29. Dorbala S, DrNayak R. Integrated data mining approach
for security alert system. International Journal of Research
Computer Communication Technology. 2014; 3(4):464–8.
30. Phua C, Lee V, Smith K, Gayler R. A comprehensive survey
of data mining-based fraud detection research. Artificial
Intelligence Computer Science. 2010; 14.
31. Sangsuriyun S, Marukatat S, Waiyamai K. Hierarchical
Multi-label Associative Classification (HMAC) using negative rules. 9th IEEE International Conference on Cognitive
Informatics (ICCI); Beijing. 2010. p. 919–24.
32. Li X, Qin D, Yu C. ACCF: Associative Classification based
on Closed Frequent itemsets. Fifth International Conference on Fuzzy Systems and Knowledge Discovery FSKD’08;
Shandong. 2008. p. 380–4.
33. Chen G, Liu H, Yu L, Wei Q, Zhang X. A new approach
to classification based on association rule mining. Decision
Support Systems. 2008; 42(2):674–89.
34. Kundu G. Islam MM. Munir S. Bari MF. ACN: An associative classifier with negative rules. 11th IEEE International
Conference on Computational Science and Engineering,
CSE’08; Sao Paulo. 2008. p. 369–75.
Indian Journal of Science and Technology
9
×

Report this document