Exercises
Exercises
Conceptual
Exercise 1 Conceptual
We fit a logistic regression to predict whether a student passes a course, from \(X_1\) the hours studied and \(X_2\) the grade in a statistics class. We obtain \(\hat\beta_0 = -6\), \(\hat\beta_1 = 0.025\) and \(\hat\beta_2 = 1\).
- Estimate the probability that a student who studies 75 hours and had a 4 in the statistics class passes.
- How many hours would that student need to study to have a 50 percent chance of passing.
- Write \(\hat\beta_1\) as an odds ratio and say in one sentence what it means.
Exercise 2 Conceptual
Here we look at the loss that maximum likelihood gives us for classification with \(C\) classes.
Write the \(i\)-th component of the linear function applied to the \(k\)-th point as \(f_i(x_k) = \theta_{i0} + \theta_{i1}x_{k1} + \cdots + \theta_{ip}x_{kp}\), so that \(\theta\) is a \(C \times (p+1)\) matrix.
- Write the log-likelihood for \(n\) points, \(p\) input dimensions and \(C\) classes, and simplify it as much as you can. Use the convention \(s_y(f(x)) = P(y|x)\) for the \(y\)-th component of the softmax.
- Take \(C = 3\) and \(p = 2\). Write the log-likelihood explicitly for the training set \(\mathcal D = ((x_1 = (0, 0), y_1 = C), (x_2 = (3, 0), y_2 = A), (x_3 = (0, 2), y_3 = B))\).
- With \(C = 2\) the softmax has twice as many parameters as logistic regression. Set \(\theta_{20} = \theta_{21} = 0\) and show that we recover logistic regression. Hint, show that \(s_1(f(x)) = s(f_1(x))\) and \(s_2(f(x)) = 1 - s(f_1(x))\).
- Show that we can always set the parameters of the last row to zero, for any \(C\). Hint, show that the softmax with \(\tilde\theta_{ij} = \theta_{ij} - \theta_{Cj}\) gives the same value.
Exercise 3 Conceptual
A test has sensitivity 0.99 and specificity 0.95. The disease affects 1 person in 1000.
- Of 100000 people, how many test positive.
- Of those, how many actually have the disease.
- The AUC of this test is high. Explain why that is not enough to make it useful for screening, and what you would change.
Applied
Exercise 4 Applied
Linear classification of the MNIST digits.
- Load MNIST. Use the first 5000 rows so that it does not take too long.
- Scale the pixel values to \([0, 1)\) by dividing by 255.
- Fit a multinomial logistic regression without penalty. Be patient.
- Compute the misclassification rate and the confusion matrix on the training set.
- Use rows 60001 to 70000 as a test set and do the same there.
- Plot some correctly classified test images and some wrongly classified ones. Are the wrong ones also hard for you.
Exercise 5 Applied
Write a data generator that samples \(x\) from a normal distribution with mean 2 and standard deviation 3, and samples \(y \in \{\text{true}, \text{false}\}\) from a Bernoulli distribution with rate \(s(0.5x - 2.7)\).
- Create a training set of size \(n = 20\) and fit a logistic regression.
- Look at the fitted parameters and compare them to the generator.
- Predict the probability of
trueon the training input. - Create a test set of size \(n = 10^4\) where the input is always \(x = 4\). Estimate the average test error at \(x = 4\) with the negative log-likelihood.
- Compute the same test error directly from the fitted parameters and the known generator, and compare.
- Rerun with different training sets of size 20, then with size \(10^4\). Write down what you observe.
Exercise 6 Applied · optional
Poisson regression on the bicycle rental data.
- Fit a Poisson regression of
countontempandhumidity. - Fit a linear regression on the same inputs and output.
- Make a scatter plot with the true counts on the horizontal axis and the predictions of both models on the vertical axis. Add the diagonal.
- Comment on the differences you see.