Practical Business Python

Taking care of business, one python script at a time

Tue 02 June 2020

sidetable - Create Simple Summary Tables in Pandas

Posted by Chris Moffitt in articles   

article header image

Introduction

Today I am happy to announce the release of a new pandas utility library called sidetable. This library makes it easy to build a frequency table and simple summary of missing values in a DataFrame. I have found it to be a useful tool when starting data exploration on a new data set and I hope others find it useful as well.

This project is also an opportunity to illustrate how to use pandas new API to register custom DataFrame accessors. This API allows you to build custom functions for working with pandas DataFrames and Series and could be really useful for building out your own library of custom pandas accessor functions.

sidetable

At its core, sidetable is a super-charged version of pandas value_counts with a little bit of crosstab mixed in. For instance, let’s look at some data on School Improvement Grants so we can see how sidetable can help us explore a new data set and figure out approaches for more complex analysis.

The only external dependency is pandas version >= 1.0. Make sure it is installed, then install sidetable:

python -m pip install sidetable

Once sidetable is installed, you need to import it to get the pandas accessor registered.

import pandas as pd
import sidetable

df = pd.read_csv('https://github.com/chris1610/pbpython/blob/master/data/school_transform.csv?raw=True', index_col=0)
Data from file

Now that sidetable is imported, you have a new accessor on all your DataFrames - stb that you can use to build summary tables. For instance, we can use .stb.freq() to build a frequency table to show how many schools were included by state with cumulative totals and percentages:

df.stb.freq(['State'])
State frequency table

This example shows that CA occurs 92 times and represents 12.15% of the total number of schools. If you include FL in the counts, you now have 163 total schools that represent 21.5% of the total.

As of version 0.6, sidetable displays percents correctly. The examples shown here are proportions not percentages. Newer versions will multiply the results times 100 - which is the correct way to represent a percent. Recent versions of sidetable also use lower cases titles for column names. Users are encouraged to use the most recent version.

For comparison, here’s value_counts(normalize=True) next to sidetable’s output:

df.info()

I think you’ll agree sidetable provides a lot more insight with not much more effort.

But wait, there’s more!

What if we want a quick view of the states that contribute around 50% of the total? Use the thresh argument to group all of the rest into an “Others” category:

df.stb.freq(['State'], thresh=.5)
Top 50 percent

This is handy. Now we can see that 8 states contributed almost 50% of the total and all the other states account for the remainder.

If we want, we can rename the catch-all category using other_label

df.stb.freq(['State'], thresh=.5, other_label='Rest of states')

One of the useful features of sidetable is that it can group columns together to further understand the distribution. For instance, what if we want to see how the various “Transformation Models” are applied across Regions?

df.stb.freq(['Region', 'Model Selected'])
Region and Model Selected

This view is a quick way to understand the interaction and distribution of the various data elements. I find that this is an easy way to explore data and get some insights that might warrant further analysis. A table like this is also easy to share with others since it is relatively simple to understand.

You could definitely perform this analysis with standard pandas (that’s all that is behind the scenes after all). It is cumbersome though, to remember the code. My experience is that if it is tough to remember then you are less likely to do it. simpletable tries to make this type of summary very easy to do.

Up until now, we have been counting the number of instances. What might be much more interesting is looking at the total breakdown by Award Amount . sidetable allows you to pass a value column that can be summed (instead of counting occurrences).

df.stb.freq(['Region'], value='Award_Amount')
Award distribution

This view gives us insight that the Northeast has the least amount of dollars spent on these projects and that 37% of the total spend went to schools in the South region.

Finally, we can look at the types of models selected and determine the 80/20 breakdown of the allocated dollars:

df.stb.freq(['Region', 'Model Selected'],
             value='Award_Amount', thresh=.82,
             other_label='Remaining')
Award distribution

If you’re familiar with pandas crosstab, then one way to look at sidetable is that it is an expanded version of a crosstab with some convenience functions to view the data more easily:

Cross tab vs. sidetable

One of sidetable’s goals is that its output is easy to interpret. If you would like to leverage pandas style functions to format your output for improved readability, sidetable can format Percentage and Amount columns to be more readable. This is not used by default but can be seen by passing style=True to the function:

df.stb.freq(['Region'], value='Award_Amount', style=True)
Formatted tables

So far I have only shown the freq function but in the interest of showing how to add other functions to the library, here’s an example of building a simple missing values table:

df.stb.missing()
Missing values

In this table, there are 10 missing values in the Region column that represent a little less than 1.3% of the total values in that column.

You can get similar information using df.info() but I find this table easier to interpret when it comes to quickly identifying missing values:

df.info()

The documentation shows more information on usage and other options. Please check it out and let me know if it is useful to you.

One thing I do want to do is thank three people for their contributions to make sidetable work.

  • Peter Baumgartner - For the original inspiration in this tweet thread
  • Steve Miller - For an article that illustrates the value of looking at frequency distribution article
  • Ted Petrou - Made this post showing how to count null values in a DataFrame.

Each of these references was leveraged very heavily to make sidetable. Thank you!

Finally, the functionality in missing is not meant to be a replacement for the excellent missingno module. The implementation included in sidetable is a quick summary version and does not include any of the useful visualizations in missingno.

Introducing the pandas accessor API

If you would like to learn how to build your own accessor, it’s actually relatively straightforward. As a reference, you can view the file that does all the work here.

Here’s a short summary of how to get started. At the top of your file import pandas to get access to the decorator:

import pandas as pd

@pd.api.extensions.register_dataframe_accessor("stb")
class SideTableAccessor:

    def __init__(self, pandas_obj):
        self._validate(pandas_obj)
        self._obj = pandas_obj

This portion of code creates the accessor class and defines the accessor value which I have chosen as stb . Once this is in place, any time you import the python module containing this code, you will get the accessor registered and available on all DataFrames.

When the class is instantiated, the current pandas DataFrame will be validated through the _validate() method and then the DataFrame will be reference in subsequent functions using self._obj

In this case, I don’t really do much with the validate method but you could choose to add more logic:

@staticmethod
def _validate(obj):
    # verify this is a DataFrame
    if not isinstance(obj, pd.DataFrame):
        raise AttributeError("Must be a pandas DataFrame")

All of the work is done in the freq and missing functions. For the most part, it is all standard pandas code. You just need to make sure you return a valid DataFrame.

For example, here is the full version of the missing function at the time of this article:

def missing(self, clip_0=False, style=False):
    """ Build table of missing data in each column.

        clip_0 (bool):     In cases where 0 counts are generated, remove them from the list
        style (bool):     Apply a pandas style to format percentages

    Returns:
        DataFrame with each Column including total Missing Values, Percent Missing
        and Total rows
    """
    missing = pd.concat([self._obj.isna().sum(),
                         self._obj.isna().mean()],
                        axis='columns').rename(columns={
                            0: 'Missing',
                            1: 'Percent'
                        })
    missing['Total'] = len(self._obj)
    if clip_0:
        missing = missing[missing['Missing'] > 0]

    results = missing[['Missing', 'Total',
                       'Percent']].sort_values(by=['Missing'],
                                               ascending=False)
    if style:
        format_dict = {'Percent': '{:.2%}', 'Total': '{0:,.0f}'}
        return results.style.format(format_dict)
    else:
        return results

In your “normal” pandas code, you would reference the DataFrame using df but here, use self._obj as your DataFrame to perform your concatenation and sorting.

I can see this as a very useful approach for building your own custom flavor of pandas functions. If you have certain transformation, cleaning or summarizing data that you do, then this might be an approach to consider - instead of just copying and pasting the code from file to file.

Summary

Pandas has a very rich API but sometimes it can take a lot of typing and wrangling to get the data in the format that is easy to understand. sidetable can make some of those summary tasks a lot easier by building frequency tables on combinations of your data and identifying gaps in your data.

sidetable does not replace any of the sophisticated analysis you will likely need to do to answer complex questions. However, it is a handy tool for quickly analyzing your data and identifying patterns you may want to investigate further.

In addition, I want sidetable to serve as an example of how to build you own pandas accessor that streamlines your normal analysis process.

I hope you find sidetable useful. If you have ideas for improvements or bug reports, head on over to github and let me know. I hope this can grow over time and become a useful tool that helps many others. I am curious to see what the community does with it.

Comments