The k-Nearest Neighbors algorithm (or kNN for short) is an easy algorithm to understand and to implement, and a powerful tool to have at your disposal.

In this tutorial you will implement the k-Nearest Neighbors algorithm from scratch in Python (2.7). The implementation will be specific for classification problems and will be demonstrated using the Iris flowers classification problem.

This tutorial is for you if you are a Python programmer, or a programmer who can pick-up python quickly, and you are interested in how to implement the k-Nearest Neighbors algorithm from scratch.

## What is k-Nearest Neighbors

The model for kNN is the entire training dataset. When a prediction is required for a unseen data instance, the kNN algorithm will search through the training dataset for the k-most similar instances. The prediction attribute of the most similar instances is summarized and returned as the prediction for the unseen instance.

The similarity measure is dependent on the type of data. For real-valued data, the Euclidean distance can be used. Other other types of data such as categorical or binary data, Hamming distance can be used.

In the case of regression problems, the average of the predicted attribute may be returned. In the case of classification, the most prevalent class may be returned.

## How does k-Nearest Neighbors Work

The kNN algorithm is belongs to the family of instance-based, competitive learning and lazy learning algorithms.

Instance-based algorithms are those algorithms that model the problem using data instances (or rows) in order to make predictive decisions. The kNN algorithm is an extreme form of instance-based methods because all training observations are retained as part of the model.

It is a competitive learning algorithm, because it internally uses competition between model elements (data instances) in order to make a predictive decision. The objective similarity measure between data instances causes each data instance to compete to “win” or be most similar to a given unseen data instance and contribute to a prediction.

Lazy learning refers to the fact that the algorithm does not build a model until the time that a prediction is required. It is lazy because it only does work at the last second. This has the benefit of only including data relevant to the unseen data, called a localized model. A disadvantage is that it can be computationally expensive to repeat the same or similar searches over larger training datasets.

Finally, kNN is powerful because it does not assume anything about the data, other than a distance measure can be calculated consistently between any two instances. As such, it is called non-parametric or non-linear as it does not assume a functional form.

## Get your FREE Algorithms Mind Map

I've created a handy mind map of 60+ algorithms organized by type.

Download it, print it and use it.

Also get exclusive access to the machine learning algorithms email mini-course.

## Classify Flowers Using Measurements

The test problem we will be using in this tutorial is iris classification.

The problem is comprised of 150 observations of iris flowers from three different species. There are 4 measurements of given flowers: sepal length, sepal width, petal length and petal width, all in the same unit of centimeters. The predicted attribute is the species, which is one of setosa, versicolor or virginica.

It is a standard dataset where the species is known for all instances. As such we can split the data into training and test datasets and use the results to evaluate our algorithm implementation. Good classification accuracy on this problem is above 90% correct, typically 96% or better.

You can download the dataset for free from iris.data, see the resources section for further details.

## How to implement k-Nearest Neighbors in Python

This tutorial is broken down into the following steps:

**Handle**Data: Open the dataset from CSV and split into test/train datasets.**Similarity**: Calculate the distance between two data instances.**Neighbors**: Locate k most similar data instances.**Response**: Generate a response from a set of data instances.**Accuracy**: Summarize the accuracy of predictions.**Main**: Tie it all together.

### 1. Handle Data

The first thing we need to do is load our data file. The data is in CSV format without a header line or any quotes. We can open the file with the open function and read the data lines using the reader function in the csv module.

1 2 3 4 5 |
import csv with open('iris.data', 'rb') as csvfile: lines = csv.reader(csvfile) for row in lines: print ', '.join(row) |

Next we need to split the data into a training dataset that kNN can use to make predictions and a test dataset that we can use to evaluate the accuracy of the model.

We first need to convert the flower measures that were loaded as strings into numbers that we can work with. Next we need to split the data set randomly into train and datasets. A ratio of 67/33 for train/test is a standard ratio used.

Pulling it all together, we can define a function called **loadDataset** that loads a CSV with the provided filename and splits it randomly into train and test datasets using the provided split ratio.

1 2 3 4 5 6 7 8 9 10 11 12 13 |
import csv import random def loadDataset(filename, split, trainingSet=[] , testSet=[]): with open(filename, 'rb') as csvfile: lines = csv.reader(csvfile) dataset = list(lines) for x in range(len(dataset)-1): for y in range(4): dataset[x][y] = float(dataset[x][y]) if random.random() < split: trainingSet.append(dataset[x]) else: testSet.append(dataset[x]) |

Download the iris flowers dataset CSV file to the local directory. We can test this function out with our iris dataset, as follows:

1 2 3 4 5 |
trainingSet=[] testSet=[] loadDataset('iris.data', 0.66, trainingSet, testSet) print 'Train: ' + repr(len(trainingSet)) print 'Test: ' + repr(len(testSet)) |

### 2. Similarity

In order to make predictions we need to calculate the similarity between any two given data instances. This is needed so that we can locate the k most similar data instances in the training dataset for a given member of the test dataset and in turn make a prediction.

Given that all four flower measurements are numeric and have the same units, we can directly use the Euclidean distance measure. This is defined as the square root of the sum of the squared differences between the two arrays of numbers (read that again a few times and let it sink in).

Additionally, we want to control which fields to include in the distance calculation. Specifically, we only want to include the first 4 attributes. One approach is to limit the euclidean distance to a fixed length, ignoring the final dimension.

Putting all of this together we can define the **euclideanDistance** function as follows:

1 2 3 4 5 6 |
import math def euclideanDistance(instance1, instance2, length): distance = 0 for x in range(length): distance += pow((instance1[x] - instance2[x]), 2) return math.sqrt(distance) |

We can test this function with some sample data, as follows:

1 2 3 4 |
data1 = [2, 2, 2, 'a'] data2 = [4, 4, 4, 'b'] distance = euclideanDistance(data1, data2, 3) print 'Distance: ' + repr(distance) |

### 3. Neighbors

Now that we have a similarity measure, we can use it collect the k most similar instances for a given unseen instance.

This is a straight forward process of calculating the distance for all instances and selecting a subset with the smallest distance values.

Below is the **getNeighbors** function that returns k most similar neighbors from the training set for a given test instance (using the already defined **euclideanDistance** function)

1 2 3 4 5 6 7 8 9 10 11 12 |
import operator def getNeighbors(trainingSet, testInstance, k): distances = [] length = len(testInstance)-1 for x in range(len(trainingSet)): dist = euclideanDistance(testInstance, trainingSet[x], length) distances.append((trainingSet[x], dist)) distances.sort(key=operator.itemgetter(1)) neighbors = [] for x in range(k): neighbors.append(distances[x][0]) return neighbors |

We can test out this function as follows:

1 2 3 4 5 |
trainSet = [[2, 2, 2, 'a'], [4, 4, 4, 'b']] testInstance = [5, 5, 5] k = 1 neighbors = getNeighbors(trainSet, testInstance, 1) print(neighbors) |

### 4. Response

Once we have located the most similar neighbors for a test instance, the next task is to devise a predicted response based on those neighbors.

We can do this by allowing each neighbor to vote for their class attribute, and take the majority vote as the prediction.

Below provides a function for getting the majority voted response from a number of neighbors. It assumes the class is the last attribute for each neighbor.

1 2 3 4 5 6 7 8 9 10 11 |
import operator def getResponse(neighbors): classVotes = {} for x in range(len(neighbors)): response = neighbors[x][-1] if response in classVotes: classVotes[response] += 1 else: classVotes[response] = 1 sortedVotes = sorted(classVotes.iteritems(), key=operator.itemgetter(1), reverse=True) return sortedVotes[0][0] |

We can test out this function with some test neighbors, as follows:

1 2 3 |
neighbors = [[1,1,1,'a'], [2,2,2,'a'], [3,3,3,'b']] response = getResponse(neighbors) print(response) |

This approach returns one response in the case of a draw, but you could handle such cases in a specific way, such as returning no response or selecting an unbiased random response.

### 5. Accuracy

We have all of the pieces of the kNN algorithm in place. An important remaining concern is how to evaluate the accuracy of predictions.

An easy way to evaluate the accuracy of the model is to calculate a ratio of the total correct predictions out of all predictions made, called the classification accuracy.

Below is the **getAccuracy** function that sums the total correct predictions and returns the accuracy as a percentage of correct classifications.

1 2 3 4 5 6 |
def getAccuracy(testSet, predictions): correct = 0 for x in range(len(testSet)): if testSet[x][-1] is predictions[x]: correct += 1 return (correct/float(len(testSet))) * 100.0 |

We can test this function with a test dataset and predictions, as follows:

1 2 3 4 |
testSet = [[1,1,1,'a'], [2,2,2,'a'], [3,3,3,'b']] predictions = ['a', 'a', 'a'] accuracy = getAccuracy(testSet, predictions) print(accuracy) |

### 6. Main

We now have all the elements of the algorithm and we can tie them together with a main function.

Below is the complete example of implementing the kNN algorithm from scratch in Python.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 |
# Example of kNN implemented from Scratch in Python import csv import random import math import operator def loadDataset(filename, split, trainingSet=[] , testSet=[]): with open(filename, 'rb') as csvfile: lines = csv.reader(csvfile) dataset = list(lines) for x in range(len(dataset)-1): for y in range(4): dataset[x][y] = float(dataset[x][y]) if random.random() < split: trainingSet.append(dataset[x]) else: testSet.append(dataset[x]) def euclideanDistance(instance1, instance2, length): distance = 0 for x in range(length): distance += pow((instance1[x] - instance2[x]), 2) return math.sqrt(distance) def getNeighbors(trainingSet, testInstance, k): distances = [] length = len(testInstance)-1 for x in range(len(trainingSet)): dist = euclideanDistance(testInstance, trainingSet[x], length) distances.append((trainingSet[x], dist)) distances.sort(key=operator.itemgetter(1)) neighbors = [] for x in range(k): neighbors.append(distances[x][0]) return neighbors def getResponse(neighbors): classVotes = {} for x in range(len(neighbors)): response = neighbors[x][-1] if response in classVotes: classVotes[response] += 1 else: classVotes[response] = 1 sortedVotes = sorted(classVotes.iteritems(), key=operator.itemgetter(1), reverse=True) return sortedVotes[0][0] def getAccuracy(testSet, predictions): correct = 0 for x in range(len(testSet)): if testSet[x][-1] == predictions[x]: correct += 1 return (correct/float(len(testSet))) * 100.0 def main(): # prepare data trainingSet=[] testSet=[] split = 0.67 loadDataset('iris.data', split, trainingSet, testSet) print 'Train set: ' + repr(len(trainingSet)) print 'Test set: ' + repr(len(testSet)) # generate predictions predictions=[] k = 3 for x in range(len(testSet)): neighbors = getNeighbors(trainingSet, testSet[x], k) result = getResponse(neighbors) predictions.append(result) print('> predicted=' + repr(result) + ', actual=' + repr(testSet[x][-1])) accuracy = getAccuracy(testSet, predictions) print('Accuracy: ' + repr(accuracy) + '%') main() |

Running the example, you will see the results of each prediction compared to the actual class value in the test set. At the end of the run, you will see the accuracy of the model. In this case, a little over 98%.

1 2 3 4 5 6 7 |
... > predicted='Iris-virginica', actual='Iris-virginica' > predicted='Iris-virginica', actual='Iris-virginica' > predicted='Iris-virginica', actual='Iris-virginica' > predicted='Iris-virginica', actual='Iris-virginica' > predicted='Iris-virginica', actual='Iris-virginica' Accuracy: 98.0392156862745% |

## Ideas For Extensions

This section provides you with ideas for extensions that you could apply and investigate with the Python code you have implemented as part of this tutorial.

**Regression**: You could adapt the implementation to work for regression problems (predicting a real-valued attribute). The summarization of the closest instances could involve taking the mean or the median of the predicted attribute.**Normalization**: When the units of measure differ between attributes, it is possible for attributes to dominate in their contribution to the distance measure. For these types of problems, you will want to rescale all data attributes into the range 0-1 (called normalization) before calculating similarity. Update the model to support data normalization.**Alternative Distance Measure**: There are many distance measures available, and you can even develop your own domain-specific distance measures if you like. Implement an alternative distance measure, such as Manhattan distance or the vector dot product.

There are many more extensions to this algorithm you might like to explore. Two additional ideas include support for distance-weighted contribution for the k-most similar instances to the prediction and more advanced data tree-based structures for searching for similar instances.

## Resource To Learn More

This section will provide some resources that you can use to learn more about the k-Nearest Neighbors algorithm in terms of both theory of how and why it works and practical concerns for implementing it in code.

### Problem

### Code

This section links to open source implementations of kNN in popular machine learning libraries. Review these if you are considering implementing your own version of the method for operational use.

### Books

You may have one or more books on applied machine learning. This section highlights the sections or chapters in common applied books on machine learning that refer to k-Nearest Neighbors.

- Applied Predictive Modeling, pages 159 and 350.
- Data Mining: Practical Machine Learning Tools and Techniques, Third Edition (The Morgan Kaufmann Series in Data Management Systems), pages 76, 128 and 235.
- Machine Learning for Hackers, Chapter 10.
- Machine Learning in Action, Chapter 2.
- Programming Collective Intelligence: Building Smart Web 2.0 Applications, Chapters 2 and 8 and page 293.

## Tutorial Summary

In this tutorial you learned about the k-Nearest Neighbor algorithm, how it works and some metaphors that you can use to think about the algorithm and relate it to other algorithms. You implemented the kNN algorithm in Python from scratch in such a way that you understand every line of code and can adapt the implementation to explore extensions and to meet your own project needs.

Below are the 5 key learnings from this tutorial:

**k-Nearest Neighbor**: A simple algorithm to understand and implement, and a powerful non-parametric method.**Instanced-based method**: Model the problem using data instances (observations).**Competitive-learning**: Learning and predictive decisions are made by internal competition between model elements.**Lazy-learning**: A model is not constructed until it is needed in order to make a prediction.**Similarity Measure**: Calculating objective distance measures between data instances is a key feature of the algorithm.

Did you implement kNN using this tutorial? How did you go? What did you learn?

Jason –

I appreciate your step-by-step approach. Your explanation makes this material accessible for a wide audience.

Keep up the great contributions.

Thanks Damian!

A few changes for python 3

1.

print ‘Train set: ‘ + repr(len(trainingSet))

print ‘Test set: ‘ + repr(len(testSet))

print needs to be used with brackets

print (“Train set:” + repr(len(trainingSet)))

print (“Test set:”+ repr(len(testSet)))

2. iteritems() changed to items()

sortedVotes = sorted(classVotes.iteritems(), key=operator.itemgetter(1), reverse=True)

should be:

sortedVotes = sorted(classVotes.items(), key=operator.itemgetter(1), reverse=True)

Thanks for sharing.

A very interesting and clear article. I haven’t tried it out yet but will over the weekend.

Thanks.

Thanks Pete, let me know how you go.

Hey Jason, I’ve ploughed through multiple books and tutorials but your explanation helped me to finally understand what I was doing.

Looking forward to more of your tutorials.

Thanks Alan!

Hey Jason!

Thank you for awesome article!

Clear and straight forward explanation. I finaly understood the background under kNN.

p.s.

There’s some code errors in the article.

1) in getResponse it should be “return sortedVote[0]” instead sortedVotes[0][0]

2) in getAccuracy it should be “testSet[x][-1] IN predictions[x]” instead of IS.

Thanks Vadim!

I think the code is right, but perhaps I misunderstood your comments.

If you change getResponse to return sortedVote[0] you will get the class and the count. We don’t want this, we just want the class.

In getAccuracy, I am interested in an equality between the class strings (is), not a set operation (in).

Does that make sense?

Thank you very much for this example!

You’re welcome Mario.

Thank you for the post on kNN implementation..

Any pointers on normalization will be greatly appreciated ?

What if the set of features includes fields like name, age, DOB, ID ? What are good algorithms to normalize such features ?

Hey PVA, great question.

Notmalization is just the rescaling of numerical attributes between 0-1. Tools like scikit-learn can do it for you if you like, here’s a recipe: http://machinelearningmastery.com/rescaling-data-for-machine-learning-in-python-with-scikit-learn/

You can compute distances between strings using methods like edit distance, learn more here: http://en.wikipedia.org/wiki/Edit_distance

DOB – well the distance between two dates could be in days, hours or whatever makes sense in your domain.

ID might just be useful as some kind of indirect marker of “when the entry was added to the database” if you don’t have a “record create time”.

I hope this helps.

A million thanks !

I’ve had so many starting points for my ML journey, but few have been this clear.

Merci !

Glad to here it Landry!

Hi,

when i run the code it shows

ValueError: could not convert string to float: ‘sepallength’

what should i do to run the program.

please help me out as soon as early….

thanks in advance…

Hi kumaran,

I believe the example code still works just fine. If I copy-paste the code from the tutorial into a new file called knn.py and download iris.data into the same directory, the example runs fine for me using Python 2.7.

Did you modify the example in some way perhaps?

it is because the first line in your code may contain info about each columns,

convert

for x in range(len(dataset)-1):

to

for x in range(1,len(dataset)-1):

it will skip the first line and start reading the data from 2nd line

use

for x in range(1,len(dataset)):

if you skipped the last line also

Hi jabson ,

Thanks for your reply..

I am using Anaconda IDE 3.4 .

yes it works well for the iris dataset If i try to put some other dataset it shows value error because those datasets contains strings along with the integers..

example forestfire datasets.

X Y month day FFMC DMC DC ISI temp RH wind rain area

7 5 mar fri 86.2 26.2 94.3 5.1 8.2 51 6.7 0 0

7 4 oct tue 90.6 35.4 669.1 6.7 18 33 0.9 0 0

7 4 oct sat 90.6 43.7 686.9 6.7 14.6 33 1.3 0 0

8 6 mar fri 91.7 33.3 77.5 9 8.3 97 4 0.2 0

8 6 mar sun 89.3 51.3 102.2 9.6 11.4 99 1.8 0 0

Is it possible to classify these datasets also with your code??

please provide me if some other classifer code example in python…

HI KUMARAN

did you get the solution for the problem mentioned in your comment. I am also facing the same problem. Please help me or provide me the solution if you have..

Excellent article on knn. It made the concepts so clear.

Thanks sanksh!

I like how it is explained, simply and clear. Great job.

Thanks!

Great article Jason !! Crisp and Clear.

Nice artical Jason. I am a software engineer new to ML. Your step by step approach made learning easy and fun. Though Python was new to me, it became very easy since I could run small small snippet instead of try to understand the entire program in once.

Appreciate your hardwork. Keep it up.

Thanks Raju.

It’s really fantastic for me. I can’t find a better one

I also face the same problem with Kumaran. After checking, I think the problem “can’t convert string into float” is that the first row is “sepal_length” and so on. Python can’t convert it since it’s totally string. So just delete it or change the code a little.

Hi,

Many thanks for this details article. Any clue for the extension Ideas?

Thanks,

RK

Hi – I was wondering how we can have the data fed into the system without randomly shuffling as I am trying to make a prediction on the final line of data?

Do we remove:

if random.random() < split

and replace with something like:

if len(trainingSet)/len(dataset) < split

# if < 0.67 then append to the training set, otherwise append to test set

The reason I ask is that I know what data I want to predict and with this it seems that it could use the data I want to predict within the training set due to the random selection process.

I also have the same dilemma as you, I performed trial and error, right now I cant seem to make things right which code be omitted to create a prediction.

I am not a software engineer nor I have a background in computer science. I am pretty new to data science and ML as well, I just started learning Python and R but the experience is GREAT!

Thanks so much for this Jason!

This article was absolutely gorgeous. As a computational physicist grad student who has taken an interest in machine learning this was the perfect level to skim, get my hands dirty and have some fun.

Thank you so much for the article on this. I’m excited to see the rest of your site.

Thanks for the article!

I wished to write my own knn python program, and that was really helpful !

Thanks a lot for sharing this.

One thing you didn’t mention though is how you chose k=3.

To get a feeling of how sensitive is the accuracy % to k, i wrote a “screening” function that iterates over k on the training set using leave-one-out cross validation accuracy % as a ranking.

Would you have any other suggestions ?

This is really really helpful. Thanks man !!

An incredibly useful tutorial, Jason. Thank you for this.

Please could you show me how you would modify your code to work with a data set which comprises strings (i.e. text) and not numerical values?

I’m really keen to try this algorithm on text data but can’t seem to find a decent article on line.

Your help is much appreciated.

Mark

Nice tutorial! Very helpful in explaining KNN — python is so much easier to understand than the mathematical operations. One thing though — the way the range function works for Python is that the final element is not included.

In loadDataset() you have

`for x in range(len(dataset)-1):`

This should simply be:

`for x in range(len(dataset)):`

otherwise the last row of data is omitted!

this gets an index out of range..

Thank you so much

great

thank very much

That’s great! I’ve tried so many books and articles to start learning ML. Your article is the first clear one! Thank you a lot! Please, keep teaching us!)

Thanks Gleb!

Hi Jason,

Thanks for this amazing introduction! I have two questions that relate to my study on this.

First is, how is optimization implemented in this code?

Second is, what is the strength of the induction this algorithm is making as explained above, will this is be a useful induction for a thinking machine?

Thank you so much!

HI jason;

it is great tutorial it help me alot thanks for great effort but i have queastion what if i want to split the data in to randomly 100 training set and 50 test set and i want to generate in separate file with there values instead of printing total numbers? becaouse i want to test them in hugin

thank you so much!

Hi Jason,

It is a really great tutorial. Your article is so clear, but I have a problem.

When I run code, I see the right classification.

> predicted=’Iris-virginica’, actual=’Iris-virginica’

> predicted=’Iris-virginica’, actual=’Iris-virginica’

> predicted=’Iris-virginica’, actual=’Iris-virginica’

> predicted=’Iris-virginica’, actual=’Iris-virginica’

…

However, accuracy is 0%. I run accuracy test but there is no problem with code.

How can I fix the accuracy? Where do I make mistake?

Thanks for reply and your helps.

Hi, I solved this doing this:

Originaly, on the step 5, in the function getAccuracy you have:

…

for x in range(len(testSet)):

if testSet[x][-1] is predictions[x]:

correct += 1

…

The key here is in the IF statement:

if testSet[x][-1] is predictions[x]:

Change “IS” to “==” so the getAccuracy now is:

…

for x in range(len(testSet)):

if testSet[x][-1] == predictions[x]:

correct += 1

…

That solve the problem and works ok!!

I think setting the value of K plays an important role in the accuracy of the prediction. How to determine the best value of ‘K’ . Please suggest some best practices ?

Dear, How to do it for muticlass classifcation with data in excelsheet: images of digits(not handwritten) and label of that image in corresponding next column of excel ??

Your this tutorial is totally on numeric data, just gave me the idea with images.

Very clear explanation and step by step working make this very understandable. I am not sure why the list sortedVotes within the function getResponse is reversed, I thought getResponse is meant to return the most common key in the dictionary classVotes. If you reverse the list, doesn’t this return the least common key in the dictionary?

I do not know how to take the k nearest neighbour for 3 classes for ties vote for example [1,1,2,2,0]. Since for two classes, with k=odd values, we do find the maximum vote for the two classes but ties happens if we choose three classes.

Thanks in advance

hi

thanks for this great effort buddy

i have some basic questions:

1: i opened “iris.data’ file and it is simply in html window. how to download?

2: if do a copy paste technique from html page. where to copy paste?

You can use File->Save as in your browser to save the file or copy the text and paste it int a new file and save it as the file “iris.data” expected by the tutorial.

I hope that helps.

Jason.

This is a really simple but thorough explaination. Thanks for the efforts.

Could you suggest me how to draw a scatter plot for the 3 classes. It will be really great if you could upload the code. Thanks in advance!

What if we want to classify text into categories using KNN,

e.g a given paragraph of text defines {Politics,Sports,Technology}

I’m Working on a project to Classify RSS Feeds

How to download the file without using library csv at the first stage?

Nice explanation Jason.. Really appreciate your work..

Thanks Avinash.

Hi! Really comprehensive tutorial, i loved it!

What will you do if some features are more important than others to determine the right class ?

Thanks Agnes.

Often it is a good idea to perform feature selection before building your model:

http://machinelearningmastery.com/an-introduction-to-feature-selection/

Hello,

I get this error message.

Train set: 78

Test set: 21

—————————————————————————

TypeError Traceback (most recent call last)

in ()

72 print(‘Accuracy: ‘ + repr(accuracy) + ‘%’)

73

—> 74 main()

in main()

65 k = 3

66 for x in range(len(testSet)):

—> 67 neighbors = getNeighbors(trainingSet, testSet[x], k)

68 result = getResponse(neighbors)

69 predictions.append(result)

in getNeighbors(trainingSet, testInstance, k)

27 length = len(testInstance)-1

28 for x in range(len(trainingSet)):

—> 29 dist = euclideanDistance(testInstance, trainingSet[x], length)

30 distances.append((trainingSet[x], dist))

31 distances.sort(key=operator.itemgetter(1))

in euclideanDistance(instance1, instance2, length)

20 distance = 0

21 for x in range(length):

—> 22 distance += pow(float(instance1[x] – instance2[x]), 2)

23 return math.sqrt(distance)

24

TypeError: unsupported operand type(s) for -: ‘str’ and ‘str’

Can you please help.

Thank you

It is not clear, it might be a copy-paste error from the post?

Thank you for your answer,

as if i can’t do the subtraction here is the error message

TypeError: unsupported operand type(s) for -: ‘str’ and ‘str’

and i copy/past the code directly from the tutorial

am so happy to be able to extend my gratitude to you.Have searched for good books to explain machine learning(KNN) but those i came across was not as clear and simple as this brilliant and awesome step by step explanation.Indeed you are a distinguished teacher

Thanks.

hi Jason, i really want to get into Machine learning. I want to make a big project for my final year of computer engg. which i am currently in. People are really enervating that way by saying that its too far fetched for a bachelor. I want to prove them wrong. I don’t have much time (6 months from today). I really want to make something useful. Can you send me some links that can help me settle on a project with machine learning? PLZ … TYSM

import numpy as np

from sklearn import preprocessing, cross_validation, neighbors

import pandas as pd

df= np.genfromtxt(‘/home/reverse/Desktop/acs.txt’, delimiter=’,’)

X= np.array(df[:,1])

y= np.array(df[:,0])

X_train, X_test, y_train, y_test = cross_validation.train_test_split(X,y,test_size=0.2)

clf = neighbors.KNeighborsClassifier()

clf.fit(X_train, y_train)

ValueError: Found arrays with inconsistent numbers of samples: [ 1 483]

Then I tried to reshape using this code: df.reshape((483,1))

Again i am getting this error “ValueError: total size of new array must be unchanged”

Advance thanks ….

Hi Jason,

great tutorial, very easy to follow. Thanks!

One question though. You wrote:

“Additionally, we want to control which fields to include in the distance calculation. Specifically, we only want to include the first 4 attributes. One approach is to limit the euclidean distance to a fixed length, ignoring the final dimension.”

Can you explain in more detail what you mean here? Why is the final dimension ignored when we want to include all 4 attributes?

Thanks a lot,

Caroline

The gist of the paragraph is that we only want to calculate distance on input variables and exclude the output variable.

The reason is when we have new data, we will not have the output variable, only input variables. Our job will be to find the k most similar instances to the new data and discover the output variable to predict.

In the specific case, the iris dataset has 4 input variables and the 5th is the class. We only want to calculate distance using the first 4 variables.

I hope that makes things clearer.

Hi Jason! The steps u showed are great. Do you any article regarding the same in matlab.

Thank you.

Thanks Pranav,

Sorry I don’t have Matlab examples at this stage.

Best algorithm tutorial I have ever seen! Thanks a lot!

Thanks Sara, I’m glad to hear that.

Detailed explanation given and I am able to understand the algorithm/code well! Trying to implement the same with my own data set (.csv file).

loadDataset(‘knn_test.csv’, split, trainingSet, testSet)

Able to execute and get the output for small dataset (with 4-5 rows and columns in the csv file).

When I try the same code for a bigger data set with 24 columns (inputs) and 12,000 rows (samples) in the csv file, I get the following error:

TypeError: unsupported operand type(s) for -: ‘str’ and ‘str’

The following lines are indicated in the error message:

distance += pow((instance1[x] – instance2[x]), 2)

dist = euclideanDistance(testInstance, trainingSet[x], length)

neighbors = getNeighbors(trainingSet, testSet[x], k)

main()

Any help or suggestion is appreciated. Thank in advance.

Thanks Nivedita.

Perhaps the loaded data needs to be converted from strings into numeric values?

Thank you for the reply Jason. There are no strings / no-numeric values in the data set. It is a csv file with 24 columns(inputs) and 12,083 rows(samples).

Any other advice?

Help is appreciated.

Understood Nivedita, but confirm that the loaded data is stored in memory as numeric values. Print your arrays to screen and/or use type(value) on specific values in each column.

Implemented this in Golang.

Check it out at – https://github.com/vedhavyas/machine-learning/tree/master/knn

Any feedback is much appreciated.

Also planning to implement as many algorithms as possible in Golang

Well done Vedhavyas.

Thanks for your great effort and implementation but I think that you need to add normalization step before the eucledian distance calculation.

Great suggestion, thanks Baris.

In this case, all input variables have the same scale. But, I agree, normalization is an important step when the scales of the input variables different – and often even when they don’t.

Great article! It would be even fuller if you add some comments in the code; previewing the data and its structure; and a step on normalization although this dataset does not require one.

Great suggestion, thanks Sisay.

hello, i”ve some error like this:

Traceback (most recent call last):

File “C:/Users/FFA/PycharmProjects/Knn/first.py”, line 80, in

main()

File “C:/Users/FFA/PycharmProjects/Knn/first.py”, line 65, in main

loadDataset(‘iris.data’, split, trainingSet, testSet)

File “C:/Users/FFA/PycharmProjects/Knn/first.py”, line 10, in loadDataset

dataset = list(lines)

_csv.Error: iterator should return strings, not bytes (did you open the file in text mode?)

what’s wrong ? how to solve the error ?

Change this line:

to this:

See if that makes a difference.

i have the same problème, i changed previous line but it didn’t work anyway !!

how do i can plot result data set calssifier using matplotlib, thanks

Great question, sorry I don’t have an example at hand.

I would suggest using a simple 2d dataset and use a scatterplot.

hello,

iris.data site link is unreachable. Could you reupload to other site please ? Thank you

Sorry, the UCI Machine Learning Repository that hosts the datasets appears to be down at the moment.

There is a back-up for the website with all the datasets here:

http://mlr.cs.umass.edu/ml/

One of the best articles I have ever read! Everything is so perfectly explained … One BIG THANK YOU!!!

I’m so glad to hear that Gabriela.

Great tutorial, worked very well with python3 had to change the iteritems in the getResponse method to .items()

line 63 & 64:

print (“Train set: ” + repr(len(trainingSet)))

print (“Test set: ” + repr(len(testSet)))

generally great tutorial , Thank you 🙂

Thanks Abdallah.

Hi,

first of all, Thanks for this great informative tutorial.

secondly, as compared to your accuracy of ~98%, i am getting an accuracy of around ~65% for every value of k. Can you tell me if this is fine and if not what general mistake i might be doing?

Thanks 🙂

Sorry to hear that.

Perhaps a different version of Python (3 instead of 2.7?), or perhaps a copy-paste error?

Hi, Jason, this article is awesome, it really gave me clear insight of KNN, and it’s so readable. just want to thank you for your incredible work. Awesome!!

I’m glad you found it useful!

Hi,

Thanks for your article.. ?

I have something to ask you..

Is the accuracy of coding indicates the accuracy of the classification of both groups ? What if want to see the accuracy of classification of true positives ? How to coding ?

Thanks before

Yes Meaz, accuracy is on the whole problem or both groups.

You can change it to report on the accuracy of one group or another, I do not have an off the cuff snippet of code for you though.

Super Article!

After reading tones of articles in which by second paragraph I am lost, this article is like explaining Pythagoras theorem to someone who landed on Algebra!

Please keep doing this Jason

I’m glad to hear it Neeraj.

This is a great tutorial, keep it up. I am trying to use KNN to generate epsilon for my DBSCAN algorithm. My data set is a time series. It only has one feature which is sub-sequenced into different time windows. I am wondering if there is a link where I can get a clear cut explanation like this for such a problem.Do you think KNN can predict epsilon since each of my row has a unique ID not setosa etc in the iris data set.

I don’t know Afees, i would recommend try it and see.

Hi Jason

I am working on a similar solution in R but i am facing problems during training of knn

What problem are you seeing Ahmad?

Thank you very much, it really helped me to understand the concept of knn.

But when i run this clock i get an error, and i couldn’t solve it. Could you please help

import csv

import random

def loadDataset(filename, split, trainingSet=[] , testSet=[]):

with open(filename, ‘rb’) as csvfile:

lines = csv.reader(csvfile)

dataset = list(lines)

for x in range(len(dataset)):

for y in range(4):

dataset[x][y] = float(dataset[x][y])

if random.random() < split:

trainingSet.append(dataset[x])

else:

testSet.append(dataset[x])

trainingSet=[]

testSet=[]

loadDataset('iris.data', 0.66, trainingSet, testSet)

print 'Train: ' + repr(len(trainingSet))

print 'Test: ' + repr(len(testSet))

IndexError Traceback (most recent call last)

in ()

15 trainingSet=[]

16 testSet=[]

—> 17 loadDataset(‘/home/emre/SWE546_DataMining/iris’, 0.66, trainingSet, testSet)

18 print ‘Train: ‘ + repr(len(trainingSet))

19 print ‘Test: ‘ + repr(len(testSet))

in loadDataset(filename, split, trainingSet, testSet)

7 for x in range(len(dataset)):

8 for y in range(4):

—-> 9 dataset[x][y] = float(dataset[x][y])

10 if random.random() < split:

11 trainingSet.append(dataset[x])

IndexError: list index out of range

solved it thanks

Glad to hear it.

How did you solve it?

Hi jason,

I am getting error of syntax in return math.sqrt(distance) and also in undefined variables in main()

Sorry to hear that, what errors exactly?

How should I take testSet from user as input and then print my prediction as output?

AWESOME POST! I cant describe how much this has helped me understand the algorithm so I can write my own C# version. Thank you so much!

I’m glad to here!

Hello,

I have encountered a problem where I need to detect and recognize an object ( in my case a logo ) in an image. My images are some kind of scanned documents that contains mostly text, signatutes and logos. I am interested in localizing the logo and recognizing which logo is it.

My problem seems easier than most object recognition problems since the logo always comes in the same angle only the scale and position that changes. Any help on how to proceed is welcome as I’m out of options right now.

Thanks

Sound great Mark.

I expect CNNs to do well on this problem and some computer vision methods may help further.

Hi Jason, I have folowed through your tutorial and now I am trying to change it to run one of my own files instead of the iris dataset. I keep getting the error:

lines = csv.reader(csvfile)

NameError: name ‘csv’ is not defined

All i have done is change lines 62-64 from:

loadDataset(‘iris.data’, split, trainingSet, testSet)

print ‘Train set: ‘ + repr(len(trainingSet))

print ‘Test set: ‘ + repr(len(testSet))

To:

loadDataset(‘fvectors.csv’, split, trainingSet, testSet)

print( ‘Train set: ‘ + repr(len(trainingSet)))

print( ‘Test set: ‘ + repr(len(testSet)))

I have also tried to it with fvectors instead of fvectors.csv but that doesnt work either. DO you have any idea what is going wrong?

It looks like your python environment might not be installed correctly.

Consider trying this tutorial:

http://machinelearningmastery.com/setup-python-environment-machine-learning-deep-learning-anaconda/

Hi Jason, id missed an import, a silly mistake. But now i get this error:

_csv.Error: iterator should return strings, not bytes (did you open the file in text mode?)

Any ideas?

I got that fixed by changing

with open(‘fvectors.csv’, ‘rb’) as csvfile:

to

with open(‘fvectors.csv’, ‘rt’) as csvfile:

but now i get this error.

dataset[x][y] = float(dataset[x][y])

ValueError: could not convert string to float:

It appears to not like my headers or labels for the data but are the labels not essential for the predicted vs actual part of the code

Nice.

Double check you have the correct data file.

Consider opening the file in ASCII format open(filename, ‘rt’). This might work better in Python 3.

Hi Jason

thanks a lot for such a wonderful tutorial for KNN.

when i run this code i found the error as

distance += pow((instance1[x] – instance2[x]), 2)

TypeError: unsupported operand type(s) for -: ‘str’ and ‘str’

can u help me f or clearing this error

Thank u

distance += pow((float(instance1[x]) – float(instance2[x])), 2)

Hi, i have some zipcode point (Tzip) with lat/long. but these points may/maynot fall inside real zip polygon (truezip). i want to do a k nearest neighbor to see the k neighbors of a Tzip point has which majority zipcode. i mean if 3 neighbors of Tzip 77339 says 77339,77339,77152.. then majority voting will determine the class as 77339. i want Tzip and truezip as nominal variable. can i try your code for that? i am very novice at python…thanks in advance.

tweetzip, lat, long, truezip

77339, 73730.689, -990323 77339

77339, 73730.699, -990341 77339

77339, 73735.6, -990351 77152

Perhaps, you may need to tweak it for your example.

Consider using KNN from sklearn, much less code would be required:

http://machinelearningmastery.com/spot-check-classification-machine-learning-algorithms-python-scikit-learn/

Thanks for your reply. i tried to use sklearn as you suggeested. But as for line ‘kfold=model_selection.KFold(n_splits=10,random_state=seed)’ it showed an error ‘seed is not defined’.

Also i think (not sure if i am right) it also take all the variable as numeric..but i want to calculate nearest neighbor distance using 2 numeric variable (lat/long) and get result along each row.

what should i do?

def getNeighbors(trainingSet, testInstance, k):

distances = []

length = len(testInstance)-1

for x in range(len(trainingSet)):

dist = euclideanDistance(testInstance, trainingSet[x], length)

distances.append((trainingSet[x], dist))

distances.sort(key=operator.itemgetter(1))

neighbors = []

for x in range(k):

neighbors.append(distances[x][0])

return neighbors

in this fuction either “length = len(testInstance)-1” -1 shouldn’t be there or the

testInstance = [5, 5, 5] should include a character item at its last index??

Am I correct?

Yes, I believe so.

Thanks

plz anyone has dataset related to human behaviour please please share me

Consider searching kaggle and the uci machine learning repository.

Hello, can you tell me at getResponce what exactly are you doing line by line?Cause I do this in Java and cant figure out what exactly I have to do.

thanks

Hi,

I am trying to run your code in Anaconda Python —–Spyder….

I have landed in errors

(1) AttributeError: ‘dict’ object has no attribute ‘iteritems’

(2) filename = ‘iris.data.csv’

with open(filename, ‘rb’) as csvfile:

Initially while loading and opening the data file , it showed an error like

Error: iterator should return strings, not bytes (did you open the file in text mode?)

when i changed rb to rt , it works….i don’t whether it will create problem later…

Please response ASAP

Thanks

The first error may be caused because the example was developed for Python 2.7 and you are using Python 3. I hope to update the examples for Python 3 in the future.

Yes, In Python 3, change to ‘rt’ top open as a text file.

Hi, for python 3

just replace this line(47):

sortedVotes = sorted(classVotes.iteritems(), key=operator.itemgetter(1), reverse=True)

with this line:

sortedVotes = sorted(classVotes.items(), key=operator.itemgetter(1), reverse=True)

it is in def getResponse(neighbors) function

Thanks Ivan.

I didn’t find anything about performance in this article. Is it so that the performance is really bad?

let’s say we have a training set of 100,000 entries, and test set of 1000. Then the euclidean distance should be calculated 10e8 times? Any workaround for this ?

Yes, you can use more efficient distance measures (e.g. drop the sqrt) or use efficient data structures to track distances (e.g. kd-trees/balls)

Nice !! Thank you 🙂

If you are using Python 3,

Use

1.#instead of rb

with open(filename, ‘r’) as csvfile:

2. #instead of iteritems.

sortedVotes = sorted(classVotes.items(), key=operator.itemgetter(1), reverse=True)

Thanks Vipin!

Hello Jason,

Nice Article. I understand a lot about the KNN under the hood. But one thing though. In scikit learn we use KNN from training to predict in 2 step.

Step 1: Fitting the classifier

Step 2: Predicting

In Fitting section we didn’t pass the test data. Only train data is passed and hence we can see where it is training and where it is testing. With respect to your other blog on Naive Bayes implementation, the part which was calculating mean and std can be considered as fitting/training part while the part which was using Gaussian Normal Distribtuion can be considered as testing/prediction part.

However in this implementation I can not see that distinction. Can you please tell me which part should be considered as training and which part is testing. The reason I am asking this question is because it is always imporatant to correalte with scikit-learn flow so that we get a better idea.

Great question.

There is no training in knn as there is no model. The dataset is the model.

Thanks for the reply…Is it the same for even scikit learn ? What exactly happens when we fit the model for KNN in Scikit Learn then?

Yes it is the same.

Nothing I expect. Perhaps store the dataset in an efficient structure for searching (e.g. kdtree).

Thanks..That’s seems interesting..BTW..I really like your approach..Apart from your e-books what materials (video/books) you think I may need to excel in deep learning and NLP. I want to switch my career as a NLP engineer.

Practice on a lot of problems and develop real and usable skills.

Where do you think I can get best problems that would create real and usable skills? Kaggle?? or somewhere else?

See this post:

http://machinelearningmastery.com/get-started-with-kaggle/

Great post. Why aren’t you normalizing the data?

Great question. Because all features in the iris data have the same units.

HI Jason,

In one of your e-book ‘machine_learning_mastery_with_python’ Chapter – 11 (Spot-Check Classification Algorithms), you have explained KNN by using scikit learn KNeighborsClassifier class. I would like to know the difference between the detailed one what you’ve explained here and the KNeighborsClassifier class. It might be a very basic question for ML practitioner as I’m very new in ML and trying to understand the purposes of different approaches.

Thanks

Golam Sarwar

The tutorial here is to help understand how the kNN method works.

To use it in practice, I would strongly encourage you to use the implementation in a library like sklearn.

The main reasons are to avoid bugs and for performance. Learn more here:

http://machinelearningmastery.com/dont-implement-machine-learning-algorithms/

Thanks…..

nice explication and great tutorial , i hope that you have other blogs about other classification algorithms like this

thanks …. Jason

Thanks.

I do, use the blog search.

Hi Jason,

Nice explanation !!

Can you please show us the implementation of the same (KNN) algorithm in Java also ?

Thanks for the suggestion, perhaps in the future.

Thanks Jason

You’re welcome.

Hi Jason,

Is it normal to get different accuracy, FP, TP, FN, TN on every different try? I am using same data.

Yes, see this post for an explanation of why to expect this in machine learning:

https://machinelearningmastery.com/randomness-in-machine-learning/

Thanks Jason. you can add below explanation to the post to make it more clear:

I’ve discovered that the different accuracy is caused by the below line in the loadDataset function:

if random.random() randomized.csv

Hi,

I am using that function instead of getAccuracy. It gives TP, TN, FP, FN.

def getPerformance(testSet, predictions):

tp = 0

tn = 0

fp = 0

fn = 0

for x in range(len(testSet)):

if testSet[x][-1] == predictions[x]:

if predictions[x] == “yes”:

tp += 1

else:

tn += 1

else:

if predictions[x] == “yes”:

fp += 1

else:

fn += 1

performance = [ ((tp/float(len(testSet))) * 100.0), ((tn/float(len(testSet))) * 100.0), ((fp/float(len(testSet))) * 100.0), ((fn/float(len(testSet))) * 100.0) ]

return performance

HI Barrys,

What is the following line of code checking for –> if predictions[x] == ‘yes’

Seems as if it always is false….

if predictions[x] == “yes”:

tp += 1

else:

tn += 1

This is the best tutorial entry I have seen on any blog post about any topic. It is very easy to follow. The code is correct and not outdated. I love the way everything is structured. It kind of follows the TDD approach where it first builds on the production code step by step, testing each step on the way. Kudos to you for the great work! This is indeed helpful.

Thanks!

i have a probleme in reading from the dataset can you tell me wher is the problem?

import pandas as pd

import numpy as np from sklearn import preprocessing, neighbors from sklearn.model_selection import train_test_split import pandas as pd

df = np.read_txt(‘C:\Users\sms\Downloads\NSLKDD-Dataset-master\NSLKDD-Dataset-master\KDDTrain22Percent.arff’) df.replace(‘?’ , -99999, inplace=True) df.drop([‘class’], 1, inplace=True)

x = np.array(df.drop([‘class’],1)) y = np.array(df[‘class’])

x_train, x_test, y_train, y_test = train_test_split(x,y,test_size=0.2)

clf = neighbors.KNieghborsClassifier() clf.fit(x_train, y_train)

accuracy = clf.score(x_test, y_test) print(accuracy)

What is the problem?

plz upload python code for feature selection using metaheuristic firefly algorithm

Thanks for the suggestion.

Nice, I added some pre tags for you.

How can can I return no response for an unbiased random response?

I’m using this code to classify random images as letters. I have a dataset of letters for it.

For example, I have a random image that is not a letter but when I use this code to classify I get a letter in response. How can I tell that this image is not a letter? According to my dataset. Should I modify the code to check the result I get in “sortedVotes[0][1]”?

Thank you.

Perhaps you can include “non-letters” in the training dataset also?

But what if I don’t have this type of data?

Thank you.

You may have to invent or contrive it to get the results you are seeking.

Hi, I want this in java language, can you help me out with this?

You could port it to Java.

Hi, How can i plot the output of the labelled data?

What type of plot would you like?

Please do Rotation forest (with LDA and PCA) in python.

Thanks for the suggestion.

Great explanation thinking of where to start ML but this tutorial cleared my doubt and I feeling now I have been confident and can apply this algorithm to any problem thanks to you

I’m glad to hear that.

HI Jason , I seem to be getting the below error. can you please confirm whats that I need to change. quite new to python

import csv

import random

def loadDataset(filename, split, trainingSet=[] , testSet=[]):

with open(filename, ‘rt’) as csvfile:

lines = csv.reader(csvfile)

dataset = list(lines)

for x in range(len(dataset)-1):

for y in range(4):

dataset[x][y] = float(dataset[x][y])

if random.random() < split:

trainingSet.append(dataset[x])

else:

testSet.append(dataset[x])

trainingSet=[]

testSet=[]

loadDataset('iris.data',0.66, trainingSet, testSet)

print ('Train: ' + repr(len(trainingSet)))

print ('Test: ' + repr(len(testSet)))

Traceback (most recent call last):

File "”, line 17, in

loadDataset(‘iris.data’,0.66, trainingSet, testSet)

File “”, line 9, in loadDataset

dataset[x][y] = float(dataset[x][y])

ValueError: could not convert string to float: ‘5.1,3.5,1.4,0.2,Iris-setosa’

@ yuvaraj I just tried your code out (with the correct indentations) and it works perfectly for me with the given data set..

for x in range(len(dataset)-1):

for y in range(4):

dataset[x][y] = float(dataset[x][y])

These lines intend to convert dataset[x][0] dataset[x][1] dataset[x][2] dataset[x][3] from type str to type float so that they can be used for calculating the euclidean distance. You cannot convert ‘Iris-setosa’ to type float.

Hi Jason,

i’m running your code above on my dataset, it has 40’000 lines, 10 features and 1 binary class.

It takes much more time to run it (i have actually not let it finish yet, after 5-10 minutes…) compared to your 6 models code here:

https://machinelearningmastery.com/machine-learning-in-python-step-by-step/

This last code runs much much faster on the same dataset, it takes just a few seconds on a Macbook pro.

Is this normal ? Or maybe something i’m doing wrong…

I let it run today and it took about an hour, accuracy 0.96. why is the other code so much faster ? It does not run on all the data ?

It might be a hardware or environment issue?

Not sure. But you confirm the code on this page runs on all the data, not just a subset, especially the KNN code ?

https://machinelearningmastery.com/machine-learning-in-python-step-by-step/

It runs so fast.

thanks for all this work really, i’ve learned a lot.

You could try running the code on less data.

Hi, Dr. Brownlee

Perhaps instead of using

`is`

, let’s use the`==`

operator since`is`

asks for identity, and not equality. I’ve stumbled upon this error myself when trying out your tutorial. Nice work btw. Thank you!Thanks.

Thank you so much Jason. one of the best tutorial about the KNN !!

One thing..in the GetResponse function the command .iteritems() doesn’t existi anymore in Python3…instead is .items()

Thank you again

Thanks for the note.

I have trained my data using knn,with neighbours no : 3, i have calculate distane for predicted data.i got smaller and larger values as distance.

How to calculate acceptance distance for knn, how to calculate the maximum limit for distance in knn.

Please suggest any procedure to calculate maximum limit for distance in knn

Perhaps estimate these values using a test dataset.

i got very huge values a distance but it’s predicted as nearest neighbors,that is reason i wish to find the maximum acceptance distance in knn .

is there any procedure available for calculate maximum acceptance distance in knn.

There may be, I’m not across it. Perhaps check papers on google scholar.

Distance in KNN ,Please tell me what are factors will effects on distance value.

The vales of observations.