## Description

## Question 1: Density Plot (20 points)

The exercise will work on the Australia Fire dataset, which can be located here: https://www.kaggle.

com/datasets/carlosparadis/fires-from-space-australia-and-new-zeland

We will work with fire nrt V1 96617.csv which is described here: https://www.earthdata.nasa.gov/

learn/find-data/near-real-time/firms/viirs-i-band-375-m-active-fire-data

Remember that during the lecture, we emphasized several times that plotting your data is very important.

Here plot the longitude vs latitude several ways within a single figure (each in its own axes):

• Using the matplotlib defaults. (matplotlib is a python package we used in our sample code).

• Adjusting alpha and marker size to compensate for overplotting.

• Using a hexbin plot.

• Subsampling the dataset.

For each but the first one, ensure that all the plotting area is used in a reasonable way and that as much

information as possible is conveyed; this is somewhat subjective and there is no one right answer.

Answer this question based on what you find: in what areas are most of the anomalies (measurements)

located?

## Question 2: Visualizing class membership (20 points)

Visualize the distribution of Brightness temperature I-4 as a histogram (with appropriate settings). Let’s

assume we are certain of a fire if the value of temperature I-4 is saturated as visible from the histogram.

• Do a small multiples plot of whether the brightness is saturated, i.e. do one plot of lat vs long for

those points with brightness saturated and a separate for those who are not (within the same figure on

separate axes). You can pick any of the methods from the question above that you find most suitable.

Can you spot differences in the distributions?

• Plot both groups in the same axes with different colors. Try changing the order of plotting the two

classes (i.e. draw the saturated first then the non-saturated or the other way around). Make sure to

include a legend. How does that impact the result?

• Can you find a better way to compare the two distributions?

## Question 3: Regression Sydney Dataset (30 points)

You can load the Sydney dataset from https://www.kaggle.com/shree1992/housedata where you can

also find a description. The goal is to predict the ’price’ column. For this task, you can ignore the date.

• Determine which features are continuous vs. categorical. Drop rows without a valid sales price.

• Visualize the univariate distribution of each continuous variable, and the distribution of the target.

Do you notice anything? Is there something that might require special treatment?

• Visualize the dependency of the target on each continuous feature (2d scatter plot).

• Split the data in training and testing set. Use ColumnTransformer to encode categorical variables.

Impute missing values using SimpleImputer. Evaluate Linear Regression (OLS), Ridge, Lasso and

Elasticnet (although we haven’t talked about these methods yet, but you can easily find references

online and you can use provided functions by Scikit-learn or other packages directly) using crossvalidation with the default parameters. Dose scaling the data with StandardScaler help? Use the

preprocessing that works best going forward.

## Question 4: Classification on the ’credit-g’ dataset (30 points)

You can download the dataset with ’fetch openml(’credit-g’)’ and see its description at https://www.

openml.org/d/31

• Determine which features are continuous and which ones are categorical.

• Visualize the univariate distribution of each continuous variable, and the distribution of the target.

• Split the data in training and testing set. Preprocess the data (such as treatment of categorical

variables) and evaluate and initial Logistic Regression model (directly use the provided function) with

with a training/validation split.

• Use ColumnTransformer to encode categorical variables. Evaluate Logistic Regression, Linear Support

Vector Machines and nearest neighbors (You can directly call these functions). How different are the

results? How dose scaling the continuous features with StandardScaler influence the results?