Linear Classification and Nearest Neighbor Classification
1. You will use a synthetic data set for the classification task that you’ll generate yourself.
Generate two classes with 20 features each. Each class is given by a multivariate Gaussian
distribution, with both classes sharing the same covariance matrix. You are provided
with the mean vectors (DS1-m0 for mean vector of negative class and DS1-m1 for mean
vector of positive class) and the covariance matrix (DS1-cov). Generate 2000 examples
for each class, and label the data to be positive if they came from the Gaussian with
mean m1 and negative if they came from the Gaussian with mean m0. Randomly pick
30% of each class (i.e., 600 data points per class) as a test set, and train the classifiers
on the remaining 70%. data When you report performance results, it should be on the
left out 30%. Call this dataset at DS1, and submit it with your code.
2. We first consider the probabilistic LDA model as seen in class: given the class variable,
the data are assumed to be Gaussians with different means for different classes but with
the same covariance matrix. This model can formally be specified as follows:
Y ∼ Bernoulli(π), X | Y = j ∼ N (µj
Estimate the parameters of the probabilistic LDA model using the maximum likelihood
approach. For DS1, report the best fit accuracy, precision, recall and F-measure achieved
by the classifier, along with the coefficients learnt.
3. For DS1, use k-NN to learn a classifier. Repeat the experiment for different values of k
and report the performance for each value. We will compare this non-linear classifier to
the linear approach, and find out how powerful linear classifiers can be. Do you do better
than LDA or worse? Are there particular values of k which perform better? Report the
best fit accuracy, precision, recall and f-measure achieved by this classifier.
4. Now instead of having a single multivariate Gaussian distribution per class, each class
is going to be generated by a mixture of 3 Gaussians. For each class, we’ll define
3 Gaussians, with the first Gaussian of the first class sharing the covariance matrix
with the first Gaussian of the second class and so on. For both the classes, fix the
mixture probability as (0.1,0.42,0.48) i.e. the sample has arisen from first Gaussian with
probability 0.1, second with probability 0.42 and so on. Mean for three Gaussians in the
positive class are given as DS2-c1-m1, DS2-c1-m2, DS2-c1-m3. Mean for three Gaussians
in the negative class are gives as DS2-c2-m1, DS2-c2-m2, DS2-c2-m3. Corresponding 3
covariance matrices are given as DS2-cov-1, DS2-cov-2 and DS2-cov-3. Now sample from
this distribution and generate the dataset similar to question 1. Call this dataset as DS2,
and submit it with your code.
5. Now perform the experiments in questions 2 and 3 again, but now using DS2. Report
the same performance measures as before. What do you observe?
6. Comment on any similarities and differences between the performance of both classifiers
on datasets DS1 and DS2?
Instruction for code submission
1. Submit a single zipped folder with your McGill id as the name of the folder. For
example if your McGill ID is 12345678, then the submission should be 12345678.zip.
2. If you are using python, you must submit your solution as a jupyter notebook.
3. Make sure all the data files needed to run your code is within the folder and loaded with
relative path. We should be able to run your code without making any modifications.
Instruction for report submission
1. You report should be brief and to the point. When asked for comments, your comment
should not be more than 3-4 lines.
2. Do not include your code in the report!
3. If you report consists of more than one page, make sure the pages are stapled.