Scribd
Upload
40 views

Assignment 2 Solution

The K-nearest neighbor (KNN) classifier is a lazy learning algorithm that stores all available data and classifies new data based on similarity. It requires O(n) time to classify a new test sample, where n is the number of training samples. The pseudocode implements a KNN classifier on MNIST data to calculate error rates for different values of k and plots the error vs k curve.

Uploaded by

Razin
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
40 views

Assignment 2 Solution

The K-nearest neighbor (KNN) classifier is a lazy learning algorithm that stores all available data and classifies new data based on similarity. It requires O(n) time to classify a new test sample, where n is the number of training samples. The pseudocode implements a KNN classifier on MNIST data to calculate error rates for different values of k and plots the error vs k curve.

Uploaded by

Razin
Copyright
© © All Rights Reserved
Available Formats
Download as DOCX, PDF, TXT or read online on Scribd
You are on page 1/ 2

3.

K-Nearest Neighbor Classifier


3.1 Lazy Classifier
a. When a new training example becomes available, among SVM, Naive Bayes and KNN,
which classifier(s) have to be re-trained from scratch?
SVM has to be re-trained from scratch. For KNN, just add the new data to the training
set and then it will be available for prediction. Nothing else has to be done. For Naïve
Bayes also, just the count of data points will vary and accordingly probability values can
be adjusted and hence it can be updated easily. But for SVM, the new data might
change the support vectors entirely and hence has to re-trained from scratch.
b. When a new test example becomes available, among SVM, Naive Bayes and KNN, which
classifier needs the most computation to infer the class label for this example, and what
is the time complexity for this inference, assuming that we have n training examples,
and the number of features is significantly smaller than n?
KNN
In KNN, firstly we need to calculate distance from the new test sample to all of the n
training samples. Since number of features are negligible, it will take O(1) time to
calculate distance from the test sample to 1 training sample. Therefore O(n) time is
required to calculate distance from test sample to n training samples.
Now to select k closest points from the sample, need O(n Log k), assuming max-
heap is used to select k closest points.
Selecting label based on majority vote will take O(k).
Hence complexity will be O(n)+O(n Log k)+O(k). If k is negligible compared to n, then
O(n).
3.2 Implementation of KNN Classifier
a. Pseudocode
1. Download 'mnist_train.csv' and 'mnist_test.csv' files from the site mentioned.
2. Load the first 6000 samples from training set to X_train (Samples) and
y_train(Labels).
3. Load the last 1000 samples from test set to X_test (Samples) and y_test(Labels).
4. Calculate Euclidean Distance from each of the test sample to all the training samples
and store it in a matrix of 1000*6000 dimension.
5. Predict label for the test set using the distance matrix using KNN Classifier algorithm
with different values of k and calculate error.
6. Plot the graph of error vs the value of k.
Function calculate_distance_matrix
For i=0 to NumberOfTestSamples-1
difference = X_testi – X_train
squared = difference^2
summed = ∑j (squaredj)
squareRooted = √ summed
distance_matrix[i] = squareRooted
return distance_matrix

Function predict
For i=0 to NumberOfTestSamples-1
distance_from_i = distance_matrix[i]
sort(distance_from_i)
select k closest points from distance_from_i
obtain classes of those k points from y_train
y_pred[i] = majority label among k values
accuracy = (# y_pred == y_test) / (# y_pred)
error = 1-accuracy
return error
b. Curve of Error vs Value of K

You might also like