WisconsinBreastCancerDiagnosticClassificationusingKNNandRandomForest

Wisconsin Breast Cancer Diagnostic
Classification
using KNN and Random Forest
Josh Lee, Steve Ray, Stanley Ng
Abstract
Breast cancer is a disease that in which early diagnosis is extremely important as it can
facilitate the subsequent clinical management of patients. In the past a lot of people are
misdiagnosed or diagnosed for breast cancer at a later stage due to insufficient tools to
identify the patient's situation. We have taken a dataset from Dr. William H. Wolberg
who is from the University of Wisconsin Hospitals, Madison, and with this dataset, have
performed Machine learning classification method on. Therefore, we are applying
machine learning (ML) classification methods to classify breast cancer patients into
benign or malignant risk groups for this datasets by using K-Nearest Neighbor and
Random Forest classifier methods. The predictive models we are using here are based
on supervised machine learning techniques with different input attributes. In this report
we look at the statistical distribution of the data along with the different performances of
the two classifier algorithms and which of them performs better.

Contents
1. Intro
2. Data Description
3. Data Distribution and Statistical Analysis using SAS
4. K-Nearest Neighbor Classification and Results
4.1 The Basics of K-Nearest Neighbor Classification
4.2 Applying the K-Nearest Neighbor Algorithm
5. Random Forest Classification and Results
5.1 Random Forest Background
5.2 The Basics of Random Forest Algorithm
5.3 Implementing Random Forest & Experimental Results
5.4 Evaluating Random Forest Classifier Result with R
6. Conclusion
7. References

1. Introduction
In this data set, we are looking at 10 different variables and their relation to whether a
cancer cell is considered malignant or benign. In this paper, we will use several
methods of data analysis to predict how accurate our dataset is for determining which
cancer cells are malignant or benign. In the second section, we will go over the data
description, where the variables are explained in greater detail. In the third section, we
are looking at the data distribution and Statistical Analysis which will point out some of
the statistical relationships within the data. The fourth section mentions K-Nearest
Neighbor Classification, which is a method that splits the data and uses Euclidean
Distance to assess how accurate of a predictor the data is. The fifth section is Random
Forest Classification which also splits the data and uses decision trees to assess the
predictive strength of the data. The sixth section is the Conclusion, where we will sum
up our findings and provide suggestions for future projects on this topic.

2. Data Description
The data for this project was taken from Dr. William H. Wolberg who is from the
University of Wisconsin Hospitals in Madison. The data consists 700 rows of just
females each with 11 attributes, 10 of which specify the different characteristics of a
tumor that was found within the breast of each individual. This table shows what each
variable is and the values they can contain.
# Attribute Domain
-- -----------------------------------------
1. Sample code number id number
2. Clump Thickness 1 - 10
3. Uniformity of Cell Size 1 - 10
4. Uniformity of Cell Shape 1 - 10
5. Marginal Adhesion 1 - 10
6. Single Epithelial Cell Size 1 - 10
7. Bare Nuclei 1 - 10
8. Bland Chromatin 1 - 10
9. Normal Nucleoli 1 - 10
10. Mitoses 1 - 10
11. Class: (2 for benign, 4 for malignant)
We used this data to train and classify already existing observations to produce an
accuracy using two supervised machine learning algorithms, Random Forest Classifier
and K-Nearest Neighbor Classifier. We also look at each attribute and their correlation
to other existing attributes within the data set.
3. Data Distribution and Statistical Analysis using SAS
The 10 variables we are looking at in this data set all vary between 1 and 10. The
variables we found the strongest correlation between is SUnif and ShUnif with a
correlation of .90688. Some other strong correlations in the data set are those between
SUnif and BChrom, SUnif and EpSize, SUnif and BChrom, and SUnif and NNuc, all of
which have correlation coefficients between .7 and .8 . The chart below shows the more
important portion of the correlation coefficient chart.

On the next several pages, there are several scatterplots which show the data for the
variables which are most strongly correlated.

Above is the scatterplot between SUnif and ShUnif, which contain the highest
correlation between variables in the data set.
This scatterplot shows the correlation between SUnif and BChrom which has the
second strongest correlation in the data set.
Here are the density functions for shape uniformity and size uniformity, which are
extremely similar as would be expected from the strong correlation between them.

4. K-Nearest Neighbor Classification
4.1 The Basics of K-Nearest Neighbor Classification
K-Nearest Neighbor is a machine learning classification algorithm. It uses previous
observations to train it’s algorithm into classifying new observations by looking at
already classified examples. How does it use these other observations to classify new
ones? One way this can be done is through the Euclidean distance formula. The
Euclidean distance is a modified version of the pythagorean theorem[2]. The
pythagorean theorem states if you know the distance of a and b, then you can find the
distance c using this formula[3].
Knowing this formula we can apply it to the points in a 2 dimensional plane. If we know
of points q and p, then using the coordinates for those points we can calculate the
distance using the euclidean distance[2]
If q1 and p1 denote the x position for each observation in the coordinate plane, we can
apply the formula above to find the distance between those two points[2]. Knowing this,
we can take the formula a step further and look at observations with n number of
variables.

This formula calculates the distance of 2 points with n number of variables in an
n-dimensional space. K-Nearest Neighbor uses the Euclidean Distance Formula to
calculate the K number of nearest neighbors, that are already classified, for a new
observation. The algorithm then takes the majority vote of the number of neighbors to
classify the observation. Regression lines can show where the classification splits when
the number of neighbors changes to a certain amount. To visually represent this we can
look at an example. This is a 2-dimensional dataset, meaning it consists of 2 variables,
used to calculate the Euclidean Distance.
Figure: 4.1 K-NN Classification Regions [4]
In this example the number of neighbors used to classify a new observation is 15. We
can see where the regression lines are drawn to split the data into 3 different
classifications.
4.2 Applying the K-Nearest Neighbor Algorithm

Our Data set consists of 10 different useable variables to classify a patient’s tumor as
benign (2) or malignant (4). We use the K-Nearest Neighbor library provided by R to
calculate the distances between points and to do the majority vote classification of any
new observation. To test the accuracy of our algorithm, we split up our data into 2 sets,
[1]
a training set (⅔), and a test set (⅓). We do this because we want to test how well our
K-Nearest Neighbor algorithm works on other data points which we already know the
classification for. After running K-Nearest Neighbor multiple times on the data we found
that our accuracy was the same every iteration.
K:3, Accuracy: 0.9832636
K:3, Accuracy: 0.9832636
K:3, Accuracy: 0.9832636
K: 3, Accuracy: 0.9832636
The reason for this, is every time we split the data it’s always going to split into the
same training set and test set. We need to randomize the data split to get an accurate
prediction overall. We do this by splitting the data within a loop.
The next step was to find which number of neighbors (K) fits best for our algorithm.
[1]
We know that having too many K can misclassify outlier groupings, so we need to find a
number that, overall, gives us the best accuracy. Since we know that the data is split up
into two thirds and one third, with 700 observations total, then we know that we have
around 450 observations that our test set can use to calculate the neighbors. Obviously
450 neighbors is too many so we cut that down to half since we are only doing two

classification groups. After testing the number of neighbors K could be we found the
lower results for K, the better the algorithm performed.
K: 1, Accuracy: 0.9581395
K: 2, Accuracy: 0.972093
K: 3, Accuracy: 0.972093
K: 4, Accuracy: 0.972093
K: 5, Accuracy: 0.9627907
K: 6, Accuracy: 0.9674419
K: 7, Accuracy: 0.9627907
after running rerunning the algorithm with random splits for the training set and test set
100 times, we found that only 4 neighbors are needed to optimally classify a new
observation, giving us an accuracy prediction of around 97.7%.
K: 3, Overall Accuracy: 0.9774771
5. Random Forest Classification
5.1 Random Forest Background
Random Forests is an effective machine learning (ML) algorithm that uses an ensemble
learning of decision trees. In this context, ensemble learning uses multiple learners or
hypothesis for coming up with predictions, often times it outperforms classifier that only
use one algorithm alone.
5.2 The Basics of Random Forest Algorithm
Suppose matrix S is a matrix of training sample as input to the Random Forest
algorithm to create a classification model.

In this case, fA1,fA2...fAn, are the feature A of the 1st sample, followed by feature B to
the Nth sample, and feature C to the Nth sample. Last column C1-Cn represents the
training class (a.k.a Target column). The goal is to create a random forest to classify
this sample set.
Create random subsets from the sample set:
Next, we create a lot of subsets with random values from the sample set. For example,
for S1 subset, we use the line row 12 to row 35. Then we create Decision Tree #1, for
S2 subset, we use the line row 2 to 20, forming Decision Tree #2. From this M sample,
we can create the Decision Tree #M.
Use all the decision trees to create a ranking of classifiers:
Figure 5.1 Random Forest Tree Illustration

The general thumb rules for random forest is to build a bunch of trees (a forest) and
have them vote on the prediction. From the graph above, we import S1,S2..Sn Decision
Trees to RF model to make the class prediction. Suppose Tree-1 is Class A, Tree-2 is
Class B, Tree-N is Class B, the majority voting account number of votes for each
classes and predicted the Final-Class as Class B.
Measure of Node Impurity to get the best split in Decision Tree:
In this project, we applied the default random forest GINI Index [5] as the measure of
node impurity. Other measures that can be implemented are Entropy, Misclassification
error, etc.
Reference: 2013 SciRes.com
Given a node t and estimated class probabilities p(k|t) k = 1,...Q, where Q is the number
of classes. To calculate the node impurity, at each node the decrease in the GINI index
G(t) is calculated for variable tj used to make the split, where the variable xj is used to
split a node.
5.3 Implementing Random Forest & Experimental Results
The Wisconsin Breast Cancer Diagnosis Datasets problem has two class outcomes:
“Malignant” or “Benign”. In this section we present experimental results of the Random
Forest Classification method through 20 times of trials. This is done in R program
language with RF installed package [6]. The datasets were cleaned and divided
randomly into training set and test set in the ration of ( ⅔,⅓ ).
Parameters Proposed to fit RF model:
Target Column: Class
Data: Breast Cancer Wisconsin Training Set
Number of Trees in RF:500

The RF method was then experimented 20 times. The graph figure 1. below shows the
result of 20 trials of proposed method on training set and validation set. [File: Result.xls]
Figure 5.2. Result of 20 trials of RF method on training set and
Validation set of Wisconsin Breast Cancer Dataset.
Next, we analyze how the numbers of trees in Random Forest would affect classification
error. In Figure 2. We can clearly see there’s a prominent decrease after increasing the
number of trees until 120, after that the results shows a stagnant trend. Classification
error included in this plot are: err.rate,oob.times error,confusion error.
The oob error is called the Out-of-Bag error, which is an internal error estimate of a
random forest as it is being constructed, also known as Misclassification rate.[6]

Figure 5.3. The increase in number of trees for Random Forest Classification[7]
Results in a smaller classification error.
5.4 Evaluating Random Forest Classifier Result with R
In this section, I used library package (“caret”) and (“pROC”) to further analyze the
dataset result study. Documentation for “Caret”[8] and “pROC”[9] can be found under
reference.
Analysing with Caret:
With the results above, We can now run it to produce a confusion matrix with caret;
Summary of confusion.matrix:

Figure 5.4 Confusion Matrix Summary [5]
In the confusion matrix and statistics, each column holds the reference (actual data) and within
each row is the prediction. The diagonal result represent the correct predicted instances, vice
versa.
Overall accuracy is calculated at somewhere over 96% with a very minimal p-value of 2 x
10^-16. We can conclude that the classifier is doing a reasonable job classifying the cancer
class.
Next, we can look at the sensitivity and specificity to help us measure the performance of the
classifier in correctly predicting the actual class of an item, it measures true positive and true
negative performance respectively. From the data above, we can see that our classifier correctly
predicted the cancer classes 97.30% of the time. Further, when we shouldn’t have predicted
classes we didn’t we performed a 96.20% correctly. We can further visualize this in the ROC
curve Figure 3.

Figure 5.5. ROC curves of Random Forest Classification Wisconsin Breast
Cancer Dataset result.
By comparing ROC curves[10], we want the area under ROC curve to skew towards sensitivity
(True positive) to prove it is doing an excellent predicting job. In this case, our proposed method
after ROC analyzing shows that it did a excellent prediction.
Conclusion
Dataset Classification Method Classification Accuracy
Wisconsin Breast Cancer
Dataset
K-Nearest Neighbor with
varying ntrees
95%-97%
Random Forest
Classification with varying
ntrees and GINI index
96%~98%
The classification of KNN method shown an a classification accuracy ranged from 95%
to 97%, whereas the classification of Random Forest method shown a classification

accuracy from 96% to 98%. We can conclude that random forest have a slight
advantage over K-NN when it comes to prediction accuracy.
From the conclusion table above, we can overall happy to assume that both
classification method perform very well in classifying each cases. However, K-NN
method do have a really good prediction performance in this particular dataset because
of the nature of the attributes which was arranged in scaled order (1-10). Also, we can
prove that K-NN perform better in this data since it doesn’t really have any noise[11].
Given if the datasets contains more random variables and numerical values, random
forest will be a better method to choose.
We believe this 2 classification methods will be very helpful to the physicians and
doctors for their second opinion for them before making final decision in classifying
patients’ health status. By using such efficient machine learning tools, they can make
almost very accurate decisions.
References
[1] "Nearest Neighbors Classification." Nearest Neighbors Classification —
Scikit-learn 0.18.1 Documentation. N.p., n.d. Web. 07 Dec. 2016.
url:http://scikit-learn.org/stable/auto_examples/neighbors/plot_classification.html#
sphx-glr-auto-examples-neighbors-plot-classification-py
[2] 4-1. Chapter 4 Measures of Distance between Samples: Euclidean (n.d.): n.
pag. Web.
url: http://84.89.132.1/~michael/stanford/maeb4.pdf
[3] Morris, Stephanie J. "The Pythagorean Theorem." The Pythagorean Theorem.
N.p., n.d. Web. 07 Dec. 2016.
url:http://jwilson.coe.uga.edu/emt669/student.folders/morris.stephanie/emt.669/e
ssay.1/pythagorean.html
[4]"R: K-Nearest Neighbour Classification." R: K-Nearest Neighbour Classification.
N.p., n.d. Web. 07 Dec. 2016.
url: http://stat.ethz.ch/R-manual/R-devel/library/class/html/knn.html
[5] Tobias Sing, Oliver Sander, Niko Beerenwinkel, Thomas Lengauer (May
16,2013) Package’ROCR’

url:http://rocr.bioinf.mpi-sb.mpg.de/ROCR.pdf
[6] Deepanshu Bhalla (August,2016) Random Forest Explained in Simple Terms
url:http://www.listendata.com/2014/11/random-forest-with-r.html
[7] Fortran original by Leo Breiman and Adele Cutler, R port by Andy Liaw and
Matthew Wiener.(October,2015) Package “Random Forest”.
url:https://cran.r-project.org/web/packages/randomForest/randomForest.pdf
[8] Paul Ingles (July 2012) Evaluating classifier results with R Caret
url:http://oobaloo.co.uk/evaluating-classifier-results-with-r-part-2
[9]Xavier Robin, Natacha Turck, Alexandre Hainard, Natalia Tiberti, Frédérique
Lisacek, Jean-Charles Sanchez and Markus Müller (May 2015) Package “pROC”
url:https://cran.r-project.org/web/packages/pROC/index.html
[10] Raffael Vogler (June 23 2015) Illustrated Guide to ROC and AUC
url:https://www.r-bloggers.com/illustrated-guide-to-roc-and-auc/
[11] R.Y. Wang, V.C. Storey, C.P. Firth (December 2011) Introduction to noise in
data mining
url: http://sci2s.ugr.es/noisydata

WisconsinBreastCancerDiagnosticClassificationusingKNNandRandomForest

More Related Content

What's hot

Similar to WisconsinBreastCancerDiagnosticClassificationusingKNNandRandomForest

WisconsinBreastCancerDiagnosticClassificationusingKNNandRandomForest