sidetable

sidetable is a supercharged combination of pandas value_counts plus crosstab that builds simple but useful summary tables of your pandas DataFrame. sidetable can also add subtotals to your DataFrame.

Usage is straightforward. Install and import sidetable. Then access it through the new .stb accessor on your DataFrame.

For the Titanic data: df.stb.freq(['class']) will build a frequency table like this:

You can also summarize missing values with df.stb.missing():

You can group the data and add subtotals and grand totals with stb.subtotal():

df.groupby(['sex', 'class']).agg({'fare': ['sum']}).stb.subtotal()

sidetable has several useful features:

• See total counts and their relative percentages in one table. This is roughly equivalent to combining the output of value_counts() and value_counts(normalize=True) into one table.
• Include cumulative totals and percentages to better understand your thresholds. The Pareto principle applies to many different scenarios and this function makes it easy to see how your data is cumulatively distributed.
• Aggregate multiple columns together to see frequency counts for grouped data.
• Provide a threshold point above which all data is grouped into a single bucket. This is useful for quickly identifying the areas to focus your analysis.
• Get a count of the missing values in your data.
• Count the number of unique values for each column.
• Add grand totals on any DataFrame and subtotals to any grouped DataFrame

Table of Contents:

• Quick Start
• Rationale
• Installation
• Usage
  • freq
  • counts
  • missing
  • subtotal
• Caveats
• TODO
• Contributing
• Credits

Quickstart

For the impatient:

$ python -m pip install sidetable

import sidetable
import pandas as pd

# Create your DataFrame
df = pd.read_csv(myfile.csv)

# Build a frequency table for one or more columns
df.stb.freq(['column1', 'column2'])

# See what data is missing
df.stb.missing()

# Group data and add a subtotal
df.groupby(['column1', 'column2'])['col3'].sum().stb.subtotal()

That's it.

Read on for more details and more examples of what you can do sidetable.

Rationale

The idea behind sidetable is that there are a handful of useful data analysis tasks that you might run on any data set early in the data analysis process. While each of these tasks can be done in a handful of lines of pandas code, it is a lot of typing and difficult to remember.

In addition to providing useful functionality, this project is also a test to see how to build custom accessors using some of pandas relatively new API. I am hopeful this can serve as a model for other projects whether open source or just for your own usage. Please check out the release announcement for more information about the usage and how to use this as a model for your own projects.

The solutions in sidetable are heavily based on three sources:

• This tweet thread by Peter Baumgartner
• An excellent article by Steve Miller that lays out many of the code concepts incorporated into sidetable.
• Ted Petrou's post on finding the percentage of missing values in a DataFrame.

I very much appreciate the work that all three authors did to point me in this direction.

Installation

$  python -m pip install -U sidetable

This is the preferred method to install sidetable, as it will always install the most recent stable release. sidetable requires pandas 1.0 or higher and no additional dependencies. It should run anywhere that pandas runs.

If you prefer to use conda, sidetable is available on conda-forge:

$ conda install -c conda-forge sidetable

Usage

import pandas as pd
import sidetable
import seaborn as sns

df = sns.load_dataset('titanic')

sidetable uses the pandas DataFrame accessor api to add a .stb accessor to all of your DataFrames. Once you import sidetable you are ready to go. In these examples, I will be using seaborn's Titanic dataset as an example but seaborn is not a direct dependency.

freq

If you have used value_counts() before, you have probably wished it were easier to combine the values with percentage distribution.

df['class'].value_counts()

Third     491
First     216
Second    184
Name: class, dtype: int64

df['class'].value_counts(normalize=True)

Third     0.551066
First     0.242424
Second    0.206510
Name: class, dtype: float64

Which can be done, but is messy and a lot of typing and remembering:

pd.concat([df['class'].value_counts().rename('count'), 
        df['class'].value_counts(normalize=True).mul(100).rename('percentage')], axis=1)

Using sidetable is much simpler and you get cumulative totals, percents and more flexibility:

df.stb.freq(['class'])

If you want to style the results so percentages and large numbers are easier to read, use style=True:

df.stb.freq(['class'], style=True)

In addition, you can group columns together. If we want to see the breakdown among class and sex:

df.stb.freq(['sex', 'class'])

You can use as many groupings as you would like.

By default, sidetable counts the data. However, you can specify a value argument to indicate that the data should be summed based on the data in another column. For this data set, we can see how the fares are distributed by class:

df.stb.freq(['class'], value='fare')

Another feature of sidetable is that you can specify a threshold. For many data analysis, you may want to break down into large groupings to focus on and ignore others. You can use the thresh argument to define a threshold and group all entries above that threshold into an "other" grouping:

df.stb.freq(['class', 'who'], value='fare', thresh=80)

You can further customize by specifying the label to use for all the others:

df.stb.freq(['class', 'who'], value='fare', thresh=80, other_label='All others')

counts

The counts() function shows how many unique values are in each column as well as the most and least frequent values & their total counts. This summary view can help you determine if you need to convert data to a categorical value. It can also help you understand the high level structure of your data.

df.stb.counts()

By default, all data types are included but you may use the exclude and include parameters to select specific types of columns. The syntax is the same as pandas select_dtypes

For example,

df.stb.counts(exclude='number')

missing

sidetable also includes a summary table that shows the missing values in your data by count and percentage of total missing values in a column.

df.stb.missing()

If you wish to see the results with styles applied to the Percent and Total column, use:

df.stb.missing(style=True)

Finally, you can exclude the columns that have 0 missing values using the clip_0=True parameter:

df.stb.missing(clip_0=True, style=True)

subtotal

Another useful function is the subtotal function. Trying to add a subtotal to grouped pandas data is not easy. sidetable adds a subtotal() function that makes adds a subtotal at one or more levels of a DataFrame.

The subtotal function can be applied to a simple DataFrame in order to add a Grand Total label:

df.stb.subtotal()

The real power of subtotal is being able to add it to one or more levels of your grouped data. For example, you can group the data and add a subtotal at each level:

df.groupby(['sex', 'class', 'embark_town']).agg({'fare': ['sum']}).stb.subtotal()

Which yields this view (truncated for simplicity):

By default, every level in the DataFrame will be subtotaled but you can control this behavior by using the sub_level argument. For instance, you can subtotal on sex and class by passing the argument sub_level=[1,2]

summary_table = df.groupby(['sex', 'class', 'embark_town']).agg({'fare': ['sum']})
summary_table.stb.subtotal(sub_level=[1, 2])

The subtotal function also allows the user to configure the labels and separators used in the subtotal and Grand Total by using the grand_label, sub_label, show_sep and sep arguments.

Caveats

sidetable supports grouping on any data type in a pandas DataFrame. This means that you could try something like:

df.stb.freq(['fare'])

In some cases where there are a fairly small discrete number of this may be useful. However, if you have a lot of unique values, you should bin the data first. In the example, above the data would include 248 rows and not be terribly useful.

One alternative could be:

df['fare_bin'] = pd.qcut(df['fare'], q=4, labels=['low', 'medium', 'high', 'x-high'])
df.stb.freq(['fare_bin'])

The other caveat is that null or missing values can cause data to drop out while aggregating. For instance, if we look at the deck variable, there are a lot of missing values.

df.stb.freq(['deck'])

The total cumulative count only goes up to 203 not the 891 we have seen in other examples. Future versions of sidetable may handle this differently. For now, it is up to you to decide how best to handle unknowns. For example, this version of the Titanic data set has a categorical value for deck so using fillna requires an extra step:

df['deck_fillna'] = df['deck'].cat.add_categories('UNK').fillna('UNK')
df.stb.freq(['deck_fillna'])

Another variant is that there might be certain groupings where there are no valid counts.

For instance, if we look at the deck and class:

df.stb.freq(['deck', 'class'])

There are only 11 combinations. If we want to see all - even if there are not any passengers fitting that criteria, use clip_0=False

df.stb.freq(['deck', 'class'], clip_0=False)

In many cases this might be too much data, but sometimes the fact that a combination is missing could be insightful.

The final caveat relates to subtotal. When working with the subtotal function, sidetable convert a Categorical MultiIndex to a plain index in order to easily add the subtotal labels.

TODO

• [ ] Handle NaN values more effectively
• [ ] Offer binning options for continuous variables
• [ ] Offer more options, maybe plotting?

Contributing

Contributions are welcome, and they are greatly appreciated! Every little bit helps, and credit will always be given. If you have a new idea for a simple table that we should add, please submit a ticket.

For more info please click here

Credits

This package was created with Cookiecutter and the oldani/cookiecutter-simple-pypackage project template. The code used in this package is heavily based on the posts from Peter Baumgartner, Steve Miller and Ted Petrou. Thank you!

• Cookiecutter
• oldani/cookiecutter-simple-pypackage
• Peter Baumgartner - tweet thread
• Steve Miller - article
• Ted Petrou - post