4

I am building a machine learning model for a binary classification task in Python/ Jupyter Notebook. I am currently in the "Exploratory data analysis" phase and try to create multiple plots/ graphs for my data set.

My data set consists of 20 columns (19 features and 1 labeled target). Each row in my data set represents a person. There are many categorical/ nominal features in my data set and only few numerical/ continuous ones. Unfortunately I cannot upload the real data set, so I will create a dummy one.

personID age car TARGET_happiness
1 27 ford 0
2 41 tesla 1
3 55 bmw 0
4 34 tesla 1
5 62 ford 1
6 38 ford 1
7 51 bmw 0
8 46 tesla 1
9 72 bmw 0
10 59 tesla 0
11 48 ford 0
12 51 bmw 1

My aim is to create a plot/ graph to visualize the relationship between the binary variable TARGET_happiness (meaning "is the person happy?") and the categorical variable car (meaning "which car does this person own").

The plot I've used for binary TARGET_happiness vs. continuous age is a box plot, see:

box plot for binary vs continuous

This seems fine. Now I also try to use a box plot for binary TARGET_happiness vs. categorical car:

box plot for binary vs categorical

I'm not sure if this plot is useful / appropriate. Sure, you can see that Tesla owners seem to be happier than BMW owners. But the box for Ford owners looks strange.

Which type of plot/ graph can I use to better visualize the relationship between binary and categorical data?

luke359
  • 141
  • 2
    Despite the different title, just about every idea in the above thread carries over to this case. – Nick Cox Apr 21 '21 at 08:31
  • 1
    Box plots are generally useless for binary data or ordered data with only a few distinct values. If more than 25% of values are equal to the lowest value recorded then that value is both the minimum and the lower quartile and any whisker collapses to a line of zero length you can't see. A similar story applies at the other end of the distribution with the maximum and upper quartile. If less than 25% are values equal to the minimum or to the maximum, you may see a point symbol for each but the box plot alone won't tell you if that represents one data point or (almost) a quarter of them. – Nick Cox Apr 23 '21 at 10:11
  • Thinking through the definitions shows that other weird-looking box plots can arise. For example, suppose 20% of values are 1 or 2, 60% of values are 3 and 20% of values are 4 or 5. Then 3 is at once the median and both quartiles and the box collapses to a line of zero length and it's a moot point whether your software will show it. Simple bar charts will work better. – Nick Cox Apr 23 '21 at 10:16

1 Answers1

4

It makes more sense to count your 0/1 in each of the categories, for example:

import pandas as pd
import seaborn as sns

df = pd.DataFrame({'car':['ford','tesla','bmw','tesla','ford','ford','bmw','tesla','bmw','tesla','ford','bmw'],
                  'TARGET_happiness':[0,1,0,1,1,1,0,1,0,0,0,1]})


sns.catplot(x='car',hue='TARGET_happiness',data=df,kind="count")

enter image description here

Or directly using the plot method in pandas:

pd.crosstab(df['car'],df['TARGET_happiness']).plot.bar(stacked=True)

enter image description here

StupidWolf
  • 5,077