Description
Problem #1 (15 points)
Create a 2-dimensional data set with 30 samples that has the following properties
a) Samples should belong to 2 classes (15 samples per class)
b) Using a Logistic Regression classifier, all samples from both classes can be correctly classified
c) Using a K-NN classifier, with K=3, two samples from each class will always be misclassified.
The remaining 26 can be classified correctly.
Generate a scatter plot of your data. Use a different color/symbol for each class. Indicate the 4 samples
that cannot be classified correctly using the KNN and explain the reasons.
Note: This data should be generated manually and you do not need to run any code on it
Problem #2 (15 points)
Create a 2-dimensional data set with 30 samples that has the following properties
a) Samples should belong to 2 classes (15 samples per class)
b) All samples can be correctly classified using a decision tree classifier with only 2 levels
c) The data cannot be perfectly classified using a linear classifier.
If it is not possible to generate such data, explain why. Otherwise, generate a scatter plot of your data
using a different color/symbol for each class. Indicate the samples that cannot be classified correctly
using a linear classifier.
Display your 2-level decision tree indicating the feature/threshold used at each non-leaf node and
the number of samples at each leaf node.
Note: This data should be generated manually and you do not need to run any code on it
Problem #3 (25 points)
For this problem, you need to use the built-in sklearn California Housing dataset. You can load this
data using
from sklearn.datasets import fetch_california_housing
cal_housing = fetch_california_housing()
Divide the data into training and test sets using train_test_split and random_state=38
The goal is to experiment with few regression algorithms and compare their performance on this data.
a) Build and train a LASSO Regression model. Vary the constraint parameter α and analyze the
results by identifying cases of overfitting and underfitting. Select the optimal value of α and
justify your choice.
b) Build and train a Decision Tree regression model. Vary the pruning parameter and analyze
the results by identifying cases of overfitting and underfitting. Select the optimal pruning and
justify your choice.
c) Compare the accuracy of the 2 methods and the relevant features identified by each method
and comment on the results.
Problem #4 (45 points)
For this problem, you need to use the built-in sklearn digits dataset. You can load this data using
Sklearn.datasets.load_digits (*, n_class=10,return_X_y=False, as_frame=False)
Divide the data into training and test sets using train_test_split and random_state=0
The goal is to train a Random Forest classifier and optimize its performance on this data.
a) Identify the most important parameters that affect the performance of the Random Forest
classifier and outline your experimental design (using 4-fold cross validation) to learn the optimal
values for these parameters.
b) Analyze the results of the classifier using its optimal parameters and comment on its generalization
capability.
c) Visualize and explain the relevant features identified by the Random Forest classifier.
• Create a white 8×8 image that represents the original 64 features. Map each identified
relevant feature to this 2D image and display it using a grey scale that reflects its importance
(e.g. 0 most relevant feature and 255 least relevant feature).
d) Identify one misclassified sample from each class (if they exist). Visualize each misclassified
sample as an 8×8 image, and use its nearest neighbors and the learned important features to explain
why it was misclassified.
Hint: for examples on how to read this data and visualize it, check
https://scikit-learn.org/stable/auto_examples/classification/plot_digits_classification.html#sphx-glrauto-examples-classification-plot-digits-classification-py
