ENHANCED BREAST CANCER RECOGNITION BASED ON ROTATION FOREST FEATURE SELECTION ALGORITHM

Natarajan Meghanathan, et al. (Eds): SIPM, FCST, ITCA, WSE, ACSIT, CS & IT 06, pp. 223–234, 2012.
© CS & IT-CSCP 2012 DOI : 10.5121/csit.2012.2322
ENHANCED BREAST CANCER
RECOGNITION BASED ON ROTATION
FOREST FEATURE SELECTION
ALGORITHM
1
Indrajit Mandal**, 2
Sairam.N
1,2
School of Computing
SASTRA University, India-613401
1
indrajit@cse.sastra.edu**, 2
sairam@cse.sastra.edu
Abstract
Optimization problems are dominantly being solved using Computational Intelligence. One of
the issues that can be addressed in this context is problems related to attribute subset selection
evaluation. This paper presents a computational intelligence technique for solving the
optimization problem using a proposed model called Modified Genetic Search Algorithms
(MGSA) that avoids local bad search space with merit and scaled fitness variables, detecting
and deleting bad candidate chromosomes, thereby reducing the number of individual
chromosomes from search space and subsequent iterations in next generations. This paper aims
to show that Rotation forest ensembles are useful in the feature selection method. The base
classifier is multinomial logistic regression method integrated with Haar wavelets as projection
filter and reproducing the ranks of each features with 10 fold cross validation method. It also
discusses the main findings and concludes with promising result of the proposed model. It
explores the combination of MGSA for optimization with Naïve Bayes classification. The result
obtained using proposed model MGSA is validated mathematically using Principal Component
Analysis. The goal is to improve the accuracy and quality of diagnosis of Breast cancer disease
with robust machine learning algorithms. As compared to other works in literature survey,
experimental results achieved in this paper show better results with statistical inference.
Keywords:
Computational Intelligence, attribute subset selection, Rotation forest, Haar wavelets,
Modified Genetic Search algorithm.
1. Introduction
Computational Intelligence is a branch emerged from artificial intelligence. The Genetic
Algorithm (GA) is a heuristic method of evolutionary algorithm based on the idea of survival of
the fittest individual [7, 12,18]. GA is used as optimization algorithms which imitates natural

224 Computer Science & Information Technology ( CS & IT )
evolution process. It is a stochastic process method that works on the natural selection strategy
and natural genetics. It can be applied to a variety of wide range of problems. The basic steps
involved in GA are detailed in [ 2, 3, 7, 9, 11,16 ]. The GA has gained popularity due to the
inherent advantages [15] that it provides for the computation such as it does not have much
mathematical requirements [17,19,20], evolution operators performing global search and its
adaptability in hybridizing with domain dependent heuristics.
2. Modified Genetic Search Algorithm
As GA finds solution through evolution process [18], it is not inclined to good solution but move
away from bad solutions. There is a chance that it may lead solution to a dead end. Also the
number of iterations or generations required is very high. In the proposed MGSA, the idea is to
identify and eliminate the bad chromosomes using merit and scaled variables so that the search
space is reduced. Once the search space is minimized, containing prospective chromosomes then
it leads to better optimization in search process. Here the merit and scaled variables refer to the
classification error rates of chromosomes. Bad chromosomes indicate those individuals which
lead to a dead end. Here is a brief summary of MSGA:
1)Initial population generation. The parameters of the model to be optimized are considered as
chromosomes. A randomly chosen set of population of individuals is a binary string with a fixed
length.
2)Fitness function: In each generation for which GA is run, the fitness of each individual is
determined which is close to optimality. Here in the attribute selection problem, we have defined
linear function
f `= af + b, where f `, f are scaled and raw fitness values of chromosomes and a,b are constants.
3)Detection and Elimination: Those chromosomes having greater classification error rate
determined by merit and scaled variables are determined and eliminated so that it is not allowed
to be a candidate chromosome in the next generation and thereby filtering out the bad
chromosomes and hence the number of iterations in the subsequent generations gets reduced.
4)Selection: A pair of best fit chromosomes or individuals which has least classification error
rate is selected from the residual population.
5)Crossover: It is the reproductive stage where two new individuals are crossed with probability
Pc to generate new pair of offspring.
6)Mutation: A single point alteration of bit string from zero to one and vice-versa with
probability Pm is applied to selected individuals to generate a new pair of offspring and avoids
premature convergence.
3. Experimental design and results
The experimental setup exploits the use of Modified Genetic Search algorithm to optimize a
subset of inputs used to classify patients as having either benign or malignant forms of breast

Computer Science & Information Technology ( CS & IT ) 225
cancer tumors using java programming. Benign tumors are non-progressive and harmless
whereas malignant tumors spread rapidly and very harmful. The real time data is adapted from
Breast Cancer database [1] that contains 683 instances after deleting 16 records containing one or
more missing values. Besides it contains nine numeric inputs and a target attribute class which
takes on values 2(benign) and 4(malignant). Ten attributes are Clump_Thickness,
Cell_Size_Uniformity, Cell_Shape_Uniformity, Marginal_Adhesion, Single_Epi_Cell_Size,
Bare_Nuclei, Bland_Chromatin, Normal_Nucleoli, Mitoses, Class. These attributes are used in
pathology report on fine needle aspirations to determine whether a lump in a breast could be
either malignant (cancerous) or benign (non-cancerous). The details of these parameters are
discussed in [15]. The data distribution for Class and indicates that 65% (445/683) of records
have value 2(benign) while the remaining 35% (238/683) have value 4(malignant). The detail of
the feature ranking is shown below:
3.1. Feature selection
Rotation forest is an ensembles classifier based on feature extraction. The heuristic component is
the feature extraction to subset of features and rebuilding a total feature set for each classifier.
We have used ensemble that consists of multinomial logistic regression model with a ridge
estimator as classifier [21] and Haar wavelets as projection filter. To our knowledge from
literature, this ensemble is used for first time. We have found experimentally that our ensemble
gives better result compared to the ensemble discussed by Juan [22].
The proposed ensemble is detailed below:-
Let x=[x1, . . . ,xn]T
be an instance given by n variables and X be the training sample in a
form of N x n matrix. Let vector Y=[y1,….yN] be class labels , where yj takes a value from
the set. Let D1,…..,DL be the classifiers in ensemble and F is feature set.
In ensemble learning, choosing L in advance and training classifiers in parallel is
necessary.
Follow the steps to prepare the training sample for classifier Di:
1. Split F randomly into K disjoint or intersecting subsets. To maximize degree of
diversity, disjoint subsets are chosen.
2. Let Fi,j be jth subset of features to train set of classifier Di.
Draw a bootstrap sample of objects of size 75 percent by selecting randomly subset of
classes for every such subset. Run Haar wavelet for only M features in Fi,j and the
selected subset of X. Store the coefficients of the Haar wavelets components,
aij[1],…..ai,j[Mj], each of size M x 1.
3. Arrange obtained vectors using coefficients in a sparse “rotation” matrix Ri having
dimensionality n x ∑Mj. Compute the training sample for classifier Di by rearranging the
columns of Ri. Represent the rearranged rotation matrix Ra
i (size N x n). So the training
sample for classifier Di is X Ra
i.

Table 1 Attribute ranking by Rotation forest ensemble comprising multinomial logistic regression model
with a ridge estimator with Haar wavelets as projection filter using 10 fold cross validation. The ranks are
generated using Ranker search method.
average merit average rank attribute
92.704 +- 0.479 1.3 +- 0.46 Cell_Size_Uniformity
92.275 +- 0.455 1.7 +- 0.46 Cell_Shape_Uniformity
90.351 +- 0.524 3 +- 0 Bland_Chromatin
88.968 +- 0.697 4 +- 0 Bare_Nuclei
87.554 +- 0.251 5.1 +- 0.3 Single_Epi_Cell_Size
86.663 +- 0.324 6.6 +- 0.66 Normal_Nucleoli
86.409 +- 0.442 7 +- 0.89 Marginal_Adhesion
86.123 +- 0.444 7.3 +- 0.9 Clump_Thickness
78.97 +- 0.321 9 +- 0 Mitoses
3.2 Case 1:
The Naïve Bayes [5] is applied to the dataset to classify with 10-fold cross validation [10] and
shows that it achieves a very impressive 96.3397 %( 658/683) classification accuracy. Details are
shown in table2.
A confusion matrix [4, 10] shown in table 2 contains the analytical details of classifications
where nine attributes are input to it. Its performance is evaluated based on the data in the matrix
for two class classifier. Accuracy is measured by Received Operator Characteristics (ROC) [4]
area under graph with TP as Y-axis and FP as X-axis and ranges from zero to one. With area=1
represents perfect test. For class benign ROC plot =0.99 is shown in fig1 for case 1 and
Cost/benefit analysis is shown in fig 2 where gain is 0.7. When relatively the number of negative
instances are higher than positive instances, then F-measure gives better understanding of
accuracy level and is computed as F = [(β2
+1) * P * TP] / [ β2
*P+TP ] where β varies from zero
to infinity and is used to balance weight assigned to TP and P. Higher the value of F-measure
higher is the accuracy level of classifier and its value ranges from zero to one. The following
tables 2,3 show the confusion matrix and accuracy parameters respectively.
3.3 Case 2:
In case 1 all the nine attributes are considered but in real world data irrelevant, redundant or
noisy attributes are common phenomena, which impairs the result. The learning scheme Wrapper
subset evaluation [8] with Naïve Bayes classification [8] is now integrated with Genetic Search
algorithm [5]. After filtration process only seven attributes are selected as relevant. Attributes
single_cell_size and mitosis is eliminated. The attributes of Genetic algorithm [23] that includes
a population size of n=20 chromosomes, crossover probability Pc=0.6 and mutation probability
Pm=0.033.As specified, Genetic Search algorithm creates an initial set of 20 chromosomes. Now
reclassifying the records using naïve Bayes with 10-fold cross validation [13]; however, this time
only seven attributes are input to the classifier. Selected attributes: 1,2,3,4,6,7,8. They are
Clump_Thickness, Cell_Size_Uniformity Cell_Shape_Uniformity, Marginal_Adhesion,
Bare_Nuclei, Bland_Chromatin, Normal_Nucleoli.

In table 6, every subset is a chromosome and merit is the fitness score reported by naïve Bayes,
which is equal to the corresponding classification error rate. Also, each chromosome’s scaled
fitness is shown in the scaled column where we use linear scaling technique to scale the values.
By definition, the raw fitness and scaled fitness values have the linear relationship
f `= af + b (1)
where f ` and f are the scaled and raw fitness values respectively. The constants a and b are
chosen where
fàvg = favg and f`max= K fàvg. (2)
The constant K represents the expected number of copies of the fittest individual in the
population. Thus, by computing the average fitness values from table 6, we obtain fàvg
=0.055755 and favg =0.055753. To find a, b the fitness values from last two rows in table 6 are
chosen to solve simultaneous equations:
0.05911= 0.05417a + b (3)
0.06677 =0.04392a + b (4)
Solving equations (3),(4) we get a= -0.747317and b= 0.0999592. We use equation (2) to
determine K i.e. K= f`max / fàvg =0.07006/0.056999= 1.2291.
Observe in fifth row in table 6 f`=0. The raw fitness value of 0.13324 corresponds to the largest
classification error in the population produced by chromosome {8}, and as a result, f` is mapped
to zero to avoid the possibility of producing negatively scaled fitness. Here the chromosome {8}
is detected and deleted so that it does not propagate to the next generation.
The improved classifier has accuracy level of 96.9253 %, which indicates that the second model
outperforms the first model by 0.5856 % where the input to the later model is only seven
attributes indicated in ROC area. For class benign ROC plot =0.993 is shown in fig3 and
Cost/benefit analysis is shown in fig 4 where gain is 0.7 and observe the smoothness achieved in
the graph. Moreover reducing the number of attributes in second model did not affect the gain in
cost benefit analysis. The plot matrix of selected attributes is shown in fig5. That is, the
classification accuracy has improved positively where only seven of nine attributes are specified
in the input. Though the increased accuracy is not a dramatic improvement, it shows the strength
of Modified Genetic Search algorithm is an appropriate algorithm for attribute selection process.
3.4 Case 3:
Now to validate the result obtained in case 2, Principal Component Analysis (PCA) [6] with
attribute selection [14] is applied to the data set consisting of nine attributes for the attribute
selection process. It can be noticed that the accuracy of Modified Genetic Search algorithm

integrated with Naïve Bayes is mathematically proven by comparing the results with the output
from Principal component analysis like the former result only seven attributes are selected
whereas in PCA analysis result. Correlation matrix [14] of the entire data set is depicted in table
7 which gives the correlation between all pairs of attributes. Eigen vectors shown in table 8 gives
the relationship between attributes where negative and positive values show inverse and direct
proportionality respectively. The corresponding rank of attributes is shown in table 9. Therefore
it can be inferred that the strength of Modified Genetic Search algorithm is analyzed
mathematically. The output result of PCA is shown in table 7,8,9.
Fig 1 . ROC Area= 0.99 for case 1 Fig 2 . Cost/benefit analysis for gain=0.7 in case 1 with
9 attributes.
Fig 3. ROC area=0.993 for case 2 Fig 4. Cost/benefit analysis for gain=0.7 with 7
attributes in case 2

Table2: Confusion Matrix for Case 1
Predicted
Benign Malignant
425 20 Benign Actual
5 233 Malignant
Table3: Accuracy for Case1
Accuracy
by Class
TP
Rate
FP
Rate
Precision Recall F-Measure ROC
Area
Class
0.955 0.025 0.986 0.955 0.97 0.99 Benign
0.975 0.045 0.921 0.975 0.947 0.985 Malignant
Weighted
Average
0.962 0.032 0.963 0.962 0.962 0.989
Table4 Confusion matrix for Case 2
Predicted
Benign Malignant
427 18 Benign Actual
3 235 Malignant
Table 5 Accuracy for Case 2
Accuracy
by Class
TP Rate FP Rate Precision Recall F-Measure ROC Area Class
0.96 0.021 0.988 0.96 0.974 0.993 Benign
0.979 0.04 0.928 0.979 0.953 0.991 Malignant
Weighted
Average
0.966 0.028 0.967 0.966 0.967 0.992
Table 6:Initial population characteristics for the 20 chromosomes
Merit Scaled Subset
0.05417 0.05911 4 6 7 9
0.05183 0.06086 1 2 3 4 7 9
0.03953 0.07006 1 2 3 4 6 9
0.05798 0.05627 6 7 8
0.13324 0 8
0.04012 0.06962 2 3 5 6 7 8
0.04656 0.0648 2 6 7
0.09195 0.03087 5 8
0.07174 0.04598 2
0.04539 0.06568 1 6 8 9
0.04246 0.06787 3 4 5 6 7 8
0.04158 0.06853 2 4 6 7 8
0.08433 0.03656 4 5

0.06149 0.05364 2 4 7
0.04041 0.0694 1 2 4 6 7 9
0.03953 0.07006 1 3 4 6 9
0.04392 0.06677 3 6 7 8 9
0.05564 0.05802 2 4 8
0.05417 0.05911 1 4 7 8
0.04392 0.06677 3 6 7 8 9
Table 7: Correlation matrix of data set.
1 0.64 0.65 0.49 0.52 0.59 0.56 0.54 0.35
0.64 1 0.91 0.71 0.75 0.69 0.76 0.72 0.46
0.65 0.91 1 0.68 0.72 0.71 0.74 0.72 0.44
0.49 0.71 0.68 1 0.6 0.67 0.67 0.6 0.42
0.52 0.75 0.72 0.6 1 0.58 0.62 0.63 0.48
0.59 0.69 0.71 0.67 0.58 1 0.68 0.58 0.34
0.56 0.76 0.74 0.67 0.62 0.68 1 0.67 0.34
0.54 0.72 0.72 0.6 0.63 0.58 0.67 1 0.43
0.35 0.46 0.44 0.42 0.48 0.34 0.34 0.43 1
Table 8: Eigen vectors of selected attributes.
V1 V2 V3 V4 V5 V6 V7 Attributes
0.3027 -0.1424 -0.8629 -0.1024 0.0634 -0.2742 0.014 Clump_Thickness
0.3812 -0.0482 0.0153 0.2036 -0.1369 -0.0973 -0.1995 Cell_Size_Uniformity
0.3777 -0.0848 -0.0378 0.1719 -0.1043 -0.0171 -0.1242 Cell_Shape_Uniformity
0.3327 -0.0439 0.4251 -0.4651 0.0138 -0.6797 0.1256 Marginal_Adhesion
0.3363 0.1659 0.1061 0.3925 -0.6708 0.0426 0.1763 Single_Epi_Cell_Size
0.3334 -0.2546 0.0091 -0.5347 -0.123 0.604 0.3837 Bare_Nuclei
0.3461 -0.2294 0.1954 -0.011 0.251 0.2525 -0.7048 Bland_Chromatin
0.336 0.0248 0.1255 0.4475 0.6495 0.051 0.4866 Normal_Nucleoli
0.2296 0.907 -0.0885 -0.2487 0.1275 0.1415 -0.1301 Mitoses
Table 9: Ranked attributes
Ranking Attributes
0.3456 V1
0.2593 V2
0.1994 V3
0.148 V4
0.1058 V5
0.0718 V6
0.039 V7

Fig5. Plot matrix distribution of selected attributes in MGSA (Class color red: Malignant, blue: Benign )
4. Conclusion
GA is a nature’s inspired computational methodology to solve complex system optimal problems,
and exhibit its outstanding and impressive performance especially to Nondeterministic
Polynomial of some combinatorial optimizations. In this paper, the enhanced performance
accuracy of 0.5856% is achieved from proposed Modified Genetic search algorithm over that of
traditional model, demonstrated with a breast cancer data set. Also the results are validated
mathematically using PCA. Proposed method and PCA eliminated 2 attributes from the data set.
An extensive experiment has showed that improvement in initial population is effective in the
optimization process. The proposed model MGSA addresses the local optima problem using
merit and scaled variables in which the bad individual chromosomes are detected and deleted,
reducing the search space and further iterations thereby improving efficiency. The various
feature selection techniques are categorized into major category like filtering methods, wrapper
subset evaluation and embedded models. The newly introduced method provided better accuracy

as shown in the result section. The software reliability of Computer Aided Diagnosis system
(CADx) is improved by the use of machine learning ensembles.
Although in this paper there are only nine attributes, there are still 511 possible attribute subsets
that can be given as input. If the numbers of attributes increases say hundred then there are 1.27x
1030
possible attribute subsets to choose. In such a hard situation like this MGSA may prove
helpful in determining the optimal subset. From the experimental results, it is concluded that GA
is an effective algorithm that can be exploited in complex situations. It can be used in various
applications such as optimization of weights of links in neural networks. Further the performance
of GA can be improved by parallel implementation using MPI and other techniques and can be
applied to vast field of emerging applications. Further, the methods need to be tested over larger
datasets as future work.
Conflict of interest
None
References
[1] Breast cancer dataset, compiled by Dr. William H. Wolberg, University of Wisconsin Hospitals,
Madison, WI, obtained January 8, 1991.
[2] Chen Lin, “An Adaptive Genetic Algorithm Based on Population Diversity Strategy”, Third
International Conference on Genetic and Evolutionary Computing. WGEC 2009. Page(s): 93 – 96.
[3] David E. Goldberg, “Genetic algorithms in search, optimization and machine learning. Addison-
Wesley,1989.
[4] http://www2.cs.uregina.ca/~hamilton/courses/831/notes/ confusion_ matrix/confusion_matrix.html
[5] Jiawei Han, Micheline Kamber, “Data Mining Concepts and Techniques”, Morgan Kaufmann
Publishers, 2001.
[6] Joseph F. Hair,Jr., William C.Black, Barry J. Babin, Rolph E. Anderson, Ronald L.Tatham, “
Multivariate Data Analysis”, Sixth Edition,2007, Pearson Education.
[7] Kosiński, W.; Kotowski, S.; Michalewicz, Z, “On convergence and optimality of genetic algorithms”,
2010 IEEE Congress on Evolutionary Computation (CEC), Page(s): 1 – 6.
[8] Mallik, R.K. “The Uniform Correlation Matrix and its Application to Diversity” ,2007 IEEE
Transactions on Wireless Communications, Volume: 6 , Issue: 5 , Page(s): 1619 - 1625
[9] Melanie Mitchell, An introduction to Genetic Algorithms, MIT Press, Cambridge, MA,2002;first
edition,1996.
[10] Nadav David Marom, Lior Rokach, Armin Shmilovici, ”Using the Confusion Matrix for Improving
Ensemble Classifiers”, 2010 IEEE 26-th Convention of Electrical and Electronics Engineers in Israel.

[11] Nor Ashidi Mat Isa, Esugasini Subramaniam, Mohd Yusoff Mashor and Nor Hayati Othman “Fine
Needle Aspiration Cytology Evaluation for Classifying Breast Cancer Using Artificial Neural
Network”, American Journal of Applied Sciences 4 (12): 999- 1008, ISSN 1546-9239, 2007 Science
Publications
[12] Pengfei Guo, Xuezhi Wang , Yingshi Han, “The Enhanced Genetic Algorithms for the Optimization
Design”, 2010 IEEE 3rd International Conference on Biomedical Engineering and Informatics, pages:
2990-2994.
[13] Richard R. Picard and R. Dennis Cook, “Cross-Validation of Regression Models”, Journal of the
American Statistical Association, Vol. 79, No. 387 (Sep., 1984), pp. 575-583,
URL: http://www.jstor.org/stable/2288403.
[14] Ron Kohavi, George H. John (1997),“Wrappers for feature subset selection”, Artificial Intelligence.
97(1-2):273-324.
[15] Sahu, S.S.; Panda, G.; Nanda, S.J., “Improved protein structural class prediction using genetic
algorithm and artificial immune system” , World Congress on Nature & Biologically Inspired
Computing, 2009. NaBIC 2009, Page(s): 731 – 735.
[16] Zhi-Qiang Chen, Rong-Long Wang , “An Efficient Real-coded Genetic Algorithm for Real-Parameter
Optimization”, 2010 IEEE Sixth International Conference on Natural Computation, pages: 2276-
2280.
[17] Mandal, I., Sairam, N. Accurate Prediction of Coronary Artery Disease Using Reliable Diagnosis
System (2012) Journal of Medical Systems, pp. 1-21.
[18] Mandal, I., Sairam, N. Enhanced classification performance using computational intelligence (2011)
Communications in Computer and Information Science, 204 CCIS, pp. 384-391.
[19] Mandal, I. Software reliability assessment using artificial neural network (2010) ICWET 2010 -
International Conference and Workshop on Emerging Trends in Technology 2010, Conference
Proceedings, pp. 698-699. Cited 1 time.
[20] Mandal, I.A low-power content-addressable memory (CAM) using pipelined search scheme (2010)
ICWET 2010 - International Conference and Workshop on Emerging Trends in Technology 2010,
Conference Proceedings, pp. 853-858.
[21] le Cessie, S., van Houwelingen, J.C. (1992). Ridge Estimators in Logistic Regression. Applied
Statistics. 41(1):191-201.
[22] Juan J. Rodriguez, Ludmila I. Kuncheva, Carlos J. Alonso (2006). Rotation Forest: A new classifier
ensemble method. IEEE Transactions on Pattern Analysis and Machine Intelligence. 28(10):1619-
1630.
[23] Larose, D. T. (2006) Genetic Algorithms, in Data Mining Methods and Models, John Wiley & Sons,
Inc., Hoboken, NJ, USA. doi: 10.1002/0471756482.ch6

Authors
Indrajit Mandal is working as research scholar at School of Computing,
SASTRA University, India. Based on his research work, he has received two
Gold medal awards in Computer Science & Engineering discipline from
National Design and Research forum, The Institution of Engineers (India). He
has won several prizes from IITs, NITs in technical paper presentations held at
national level. He has published research papers in international peer reviewed
journals and international conferences. His research interest includes Machine
learning, Applied statistics, Computational intelligence, Software reliability.
Sairam N is working as Professor in School of Computing, SASTRA
University, India and has teaching experience of 15 years. He has published
several research papers at national and international journals and conferences.
His research interest includes Soft Computing, Theory of Computation,
Parallel Computing and Algorithms, Data Mining.

ENHANCED BREAST CANCER RECOGNITION BASED ON ROTATION FOREST FEATURE SELECTION ALGORITHM

More Related Content

What's hot (18)

Similar to ENHANCED BREAST CANCER RECOGNITION BASED ON ROTATION FOREST FEATURE SELECTION ALGORITHM (20)

More from cscpconf (20)

Recently uploaded (20)

ENHANCED BREAST CANCER RECOGNITION BASED ON ROTATION FOREST FEATURE SELECTION ALGORITHM