Table of Content

  1. What is KNN ?
  2. Where do you use KNN ?
  3. What is K in K Nearest Neighbors ?
  4. How KNN works ?
  5. What are the distance based methods used in KNN ?
  6. Disadvantages of KNN
  7. Problem solved  on Breast Cancer Dataset using KNN
    1. Initializing libraries
    2. Importing dataset
    3. Dataset Variable Terms Explained
    4. Finding missing values
    5. Exploratory Data Analysis
    6. Sorting Outlier issue
    7. Using Standardisation
    8. Splitting of Dataset to test and train
    9. Building KNeighbors classifier using simulation of different k values
    10. Building KNeighbors classifier
    11. Predicting for train data
    12. Finding the precision, recall, f1-score ,support
    13. Finding accuracy Score for Train Data
    14. Predicting Test Data
    15. Accuracy Score for test Data
    16. Pickle model as a file for usage at later stage
  8. Conclusion and Summary



What is KNN ?

KNN is expanded as K- Nearest Neighbors, which is a Supervised machine learning algorithm, a non -parametic method that solves both classification and regression problems. It predicts data using nearest neighbors by using distance  based measures. Some distance based  methods include Euclidean, Minkowsky and Manhattan.


Where do you use KNN ?

KNN is  used for both Classification and Regression problems. KNN works well for small datasets and is best suitable for clinical datasets. For Classification problems, the output will be  the class it belongs to and in case of Regression, the output will be the average of the values of neighbors considered.


What is the K in K Nearest Neighbors ?

The letter K is the paramter which tells  the number of  nearest neighbors. To predict a particular unknown point in a classification problem or regression problem, the k-value is required, which determines the number of nearest voting neighbors to find the value for the non labelled data. The K value is always a positive integer.


How KNN works ?

KNN predicts unknown value in a very interesting way ,i.e by taking the highest votes of class of its nearest neighbors. Its like telling that a person is a doctor by taking the majority votes of  his nearest neighbors  professions  which are more of doctors than engineers. KNN algorithm finds the unkown value of data point by taking the K value of nearest neighbors and finds the distance between the points and unknown point using the given distance measure(such as Euclidean ,Minkowsky ,Manhattan  or chi-sqare) .After finding the nearest neighbours by distance ,it predicts the class of unkonwn point by taking the highest number votes of the neighbor's classes. KNN doesn't learn the given dataset but simply takes it for prediction using the nearest neighbors. Hence its called the Lazy Algorithm. Just in case if the value is an unknown color, it will take majority votes as orange if the maximum number neighbors are orange and then the unkown value will be considered as orange.

The number given  to k-value should always  be taken into account during model building  since K -value can affect  accuracy and  prediction and even overfitting and underfitting of data. Ideally a odd number is preferred so as to get a decision for every unlabelled data. 


What are the distance based methods used in in KNN ?

Distance based methods used in KNN are Euclidean Distance, Minkowsky distance, Chisquare distance, Cosine similarity measure, Hamming distance, Chebychev distance and Mahalanobis distance. It's to be noted that the values in different variables in dataset should be scaled to unit norm before applying any distance based measure. Removal of outliers is also important, since KNN works on distance based algorithm and data points lying very faraway from the rest of the data can adversly create problem during distance measurement.


Disadvantages of KNN

  1. KNN cannot be used for large datasets.
  2. KNN doesnt learn the dataset and  prediction will be always based on voting method using nearest neighbors, the prediction values wont always be correct .
  3. KNN will be very slow when using large amount of data.
  4. It is sensitive to outliers and data must be scaled to one unit.
  5. KNN requires huge memory for storage and processing of large datasets.


Problem solved  on Breast Cancer Dataset using KNN


STEP 1 : Initializing libraries

    import pandas as pd
    import numpy as np
    import seaborn as sns 
    import matplotlib.pyplot as plt

STEP 2 : Importing dataset

Click here to download the dataset.

    # use pandas read csv to ingest the downloaded data into your python environment

STEP 3 : Dataset Variable Terms Explained

    #Age of person (years)
    #BMI of person (kg/m2)
    #Glucose level  (mg/dL)-Concentration of blood sugar in Humans Average blood sugar is 70 and 126 mg/dl.
    #Insulin level (µU/mL)-Insulin is a hormone produced by cells in Pancreas to control the amount of glucose(a type of sugar)flow in the blood and its absorption by the body .Normal Fasting Insulin level is 5 and 15 µU/mL.
    #HOMA  Homeostatic model assessment 
    #Leptin (ng/mL)  It helps inhibit hunger and regulate energy balance, so the body does not trigger hunger responses when it does not need energy.
    #Adiponectin (µg/mL) Adiponectin is a protein hormone and adipokine.It is involved in regulating glucose levels and fatty acid breakdown. In humans it is encoded by the ADIPOQ gene and it is produced in primarily in adipose tissue, but also in muscle, and even in the brain.
    #Resistin (ng/mL)-Resistin increases bad cholestrol in the liver. This results in heart disease.
    #MCP-1(pg/dL) Monocyte chemoattractant protein-1 (MCP-1/CCL2) is one of the key chemokines that regulate migration and infiltration of monocytes/macrophages.

    #Classification Labels:
    #1 for Healthy controls
    #2 for Patients


Coimbra data set

Figure 1 : Coimbra data set

    #Data Set no of rows and columns

116 10


STEP 4 : Finding missing values


Check missing value in data set

Figure 2 : Check missing value in data set


STEP 5 : Exploratory Data Analysis

    #Type of Data in Each Variable

Data type of individual variables in coimbra dataset.

Figure 3 : Data type of individual variables in coimbra dataset


Is multicollinearity a problem in KNN?

KNN is a non-parametic method and muticollinearity is not part of the assumptions here .The heat map shown below shows the correlation among the independent  variables in the dataset .We are not going to consider multicollinearity as we go ahead.

    #Heatmap to find correlation

Visualization of the Correlation gridFigure 4 : Visualization of the Correlation grid


    #column Values

Index(['Age', 'BMI', 'Glucose', 'Insulin', 'HOMA', 'Leptin', 'Adiponectin',
       'Resistin', 'MCP.1', 'Classification'],


STEP 6 : Sorting Outlier issue

As we discussed earlier KNN is adversly affected by outliers since its a distance based measure.Now Finding outliers and to reduce them since they might affect prediction on increase in variability since  we are using distance based measure.


    #No outliers for age

Boxplot of ageFigure 5 : Boxplot of age


    #NO outliers for BMI

Boxplot of BMIFigure 6 : Boxplot of BMI


    #Some outliers are there for Glucose and  data is Skewed 

Boxplot of GlucoseFigure 7 : Boxplot of Glucose


    #Outliers are present in Insulin

Boxplot of InsulinFigure 8 : Boxplot of Insulin


    #lots of Outliers in Homa

Boxplot of HOMAFigure 9 : Boxplot of HOMA


    #Distribution plot of HOMA

Distribution plot of HOMAFigure 10 : Distribution plot of HOMA


    #Outliers present for Leptin

Boxplot of LeptineFigure 11 : Boxplot of Leptine


    #Outliers present for Adiponectin

Boxplot of AdiponectinFigure 12 : Boxplot of Adiponectin


    #Ouliers present for Resistin

Boxplot of ResistinFigure 13 : Boxplot of Resistin


    #Outliers present for MCP.1

Boxplot of MCPFigure 14 : Boxplot of MCP


    #Removing Outliers Since they may affect prediction for KNN (quantile method)

    insulrem=cancer[(cancer['Insulin']>lowerliminsulin)&(upperliminsulin > cancer['Insulin'])]


Boxplot of Glucose in censored Insuline outliersFigure 15 : Boxplot of Glucose in censored Insuline outliers


    glucoserem=insulrem[(insulrem['Glucose'] > lowerlimglucose)&(upperlimglucose > insulrem['Glucose'])]

Boxplot of HOMA in censored Glucose outliersFigure 16 : Boxplot of HOMA in censored Glucose outliers


    homarem=glucoserem[(glucoserem['HOMA'] > lowerlimhoma)&(upperlimhoma > glucoserem['HOMA'])]

Boxplot of Adiponectin in censored HOMA outliersFigure 17 : Boxplot of Adiponectin in censored HOMA outliers


    adirem=homarem[(homarem['Adiponectin'] > lowerlimAdiponectin)&(upperlimAdiponectin > homarem['Adiponectin'])]


Boxplot of Resistin in censored Adiponectin outliersFigure 18 : Boxplot of Resistin in censored Adiponectin outliers


    Resistinrem=adirem[(adirem['Resistin'] > lowerlimresistin)&(upperlimresistin > adirem['Resistin'])]

Boxplot of Leptin in censored Resistin outliersFigure 19 : Boxplot of Leptin in censored Resistin outliers


    leptinrem=Resistinrem[(Resistinrem['Leptin'] > lowerlimLeptin)&(upperlimLeptin > Resistinrem['Leptin'])]

Boxplot of MCP in censored Leptin outliersFigure 20 : Boxplot of MCP in censored Leptin outliers


    mcprem=leptinrem[(leptinrem['MCP.1'] > lowerlimMCP)&(upperlimMCP > leptinrem['MCP.1'])]

Boxplot of MCP in final dataFigure 21 : Boxplot of MCP in final data


    # create the features from data

    # create the target variable from data


STEP 7 : Using Standardisation to bring all values to one unit since KNN is a distanced based method 

    from sklearn.preprocessing import StandardScaler

STEP 8 : Splitting of Dataset to test and train

    from sklearn.model_selection import train_test_split

STEP 9 : Building KNeighbors classifier using simulation of different k values

    #Importing KNeighbors Classifier from sklearn
    #Finding accuracies  on TrainData and Test data with euclidean distance(by default p=2)
    from sklearn.neighbors import KNeighborsClassifier
    from sklearn.metrics import accuracy_score
    for x in range(5,10,2):
        print(f'Accuracy score for  train data and test data is {acc_train_score} and {acc_test_score} respectively for {x} neighbours')

Accuracy score for train data and test data is 1.0 and 0.7142857142857143 respectively for 5 neighbours

Accuracy score for train data and test data is 1.0 and 0.7857142857142857 respectively for 7 neighbours

Accuracy score for train data and test data is 1.0 and 0.7142857142857143 respectively for 9 neighbours

The Train Accuracies inferred from the above model building with odd number of neighbors (5,7,9) are all the same 100% and Test accuracies are each different for the test Data with their own highs and lows. For 5 neighbors, accuracy is 71.42% whereas for 7 neighbors, the accuracy is 78.57%. We have made a better choice of taking 7 neighbors for the model building. This example clearly shows that neighbors play a vital impact on voting and prediction.Hence choosing the right number of neighbors is important and neighbors should always be an odd number.


STEP 10 : Building KNeighbors classifier using Eucledian Distance and 7 neighbors as optimal


STEP 11 : Predicting for train data


STEP 12 : Finding the precision, recall, f1-score ,support 

    from sklearn.metrics import classification_report

Classification report on train data set

Figure 22 : Classification report on train data set


STEP 13 : Finding accuracy Score for Train Data




STEP 14 : Predicting Test Data


STEP 15 : Accuracy Score for test Data

    from sklearn.metrics import accuracy_score


STEP 16 : Pickle model as a file for usage at later stage 

    import pickle
    #Save our model as a pickle to a file
    pickle.dump(knn, open("my_knn_model.pickle.dat", "wb"))

    # delete the existing knn model from the environment
    del knn

    #Load the pickled object from the file
    load_knn=pickle.load(open("my_knn_model.pickle.dat", "rb"))

    # Use the loaded model to make predictions

Conclusion and Summary

As you can see that the data set taken is a small clinical dataset which  has got less than 150 values and is very much suitable for building  KNN model. There were so many outliers in  almost all the independent variables which had to be removed. We scaled the dataset before model building. Then the model was built using 6 nearest neighbours with distance measure taken as Euclidean. We got the accuracy of train and test data as 100% and 78.57% respectively. Also note that we haven't specified the seed while we were doing the split in the data set. So the train and test sets would have different instances or records. Acordingly your accurracy and precision would vary. The precision and recall for the classes stands on good score. To learn more about the subject in detail, recommend to go through these free video courses. Here we should always consider to  take the best value of K for best accuracy and avoiding overfitting or underfitting of dataset.You can insert different k values and use different distance methods to get to the best optimised model. Remember to  use KNN for small datasets and its good to use other parametric or non-parametric models for large datasets.



About the Author's:

S. Ebenezer

S. Ebenezer is a Data Science intern at Simple & Real Analytics. Having a strong passion towards DataScience, he loves contributing to open source platforms. Apart from being a creative learner, he is loves working on challenging opportunities. He holds a Bachelors degree in Mechanical Engineering.


Mohan Rai

Mohan Rai is an Alumni of IIM Bangalore , he has completed his MBA from University of Pune and Bachelor of Science (Statistics) from University of Pune. He is a Certified Data Scientist by EMC. Mohan is a learner and has been enriching his experience throughout his career by exposing himself to several opportunities in the capacity of an Advisor, Consultant and a Business Owner. He has more than 18 years’ experience in the field of Analytics and has worked as an Analytics SME on domains ranging from IT, Banking, Construction, Real Estate, Automobile, Component Manufacturing and Retail. His functional scope covers areas including Training, Research, Sales, Market Research, Sales Planning, and Market Strategy.