Empirical Probabilities

This notebook describes how to compute various “empirical probabilities”: computing marginal, joint, and conditional probabilities from available data. You’ll want to be familiar with the notes on probability before working through this notebook.

It’s important to remember that using probabilities in this way is just one way of using probabilities; an empirical probability computed using the techniques in this notebook is not what a probability is; it’s just a computation that probability can be used to describe.

This notebook uses some math macros in the course site theme which may not be available in Jupyter Notebook, so if you download the notebook file you may not be able to see all math correctly.

Setup

Let’s load our common packages:

import pandas as pd
import numpy as np
import seaborn as sns
import matplotlib.pyplot as plt

Data

We’re going to do these demonstrations with the Rent the Runway data.

Let’s load it:

rtr = pd.read_json('../data/renttherunway_final_data.json.gz', lines=True)
rtr.head()
fit user_id bust size item_id weight rating rented for review_text body type review_summary category height size age review_date
0 fit 420272 34d 2260466 137lbs 10.0 vacation An adorable romper! Belt and zipper were a lit... hourglass So many compliments! romper 5' 8" 14 28.0 April 20, 2016
1 fit 273551 34b 153475 132lbs 10.0 other I rented this dress for a photo shoot. The the... straight & narrow I felt so glamourous!!! gown 5' 6" 12 36.0 June 18, 2013
2 fit 360448 NaN 1063761 NaN 10.0 party This hugged in all the right places! It was a ... NaN It was a great time to celebrate the (almost) ... sheath 5' 4" 4 116.0 December 14, 2015
3 fit 909926 34c 126335 135lbs 8.0 formal affair I rented this for my company's black tie award... pear Dress arrived on time and in perfect condition. dress 5' 5" 8 34.0 February 12, 2014
4 fit 151944 34b 616682 145lbs 10.0 wedding I have always been petite in my upper body and... athletic Was in love with this dress !!! gown 5' 9" 12 27.0 September 26, 2016

A little exploration - how many data points do we have here?

len(rtr)
192544

What is the distribution of ‘fit’ codes?

sns.countplot(x='fit', data=rtr)
plt.title('Distribution of Fit Results')
plt.show()
../../../_images/EmpiricalProbabilities_9_0.png

What is the distribution of garment categories?

plt.figure(figsize=(7,12))
sns.countplot(y='category', data=rtr)
plt.title('Distribution of Categories')
plt.show()
../../../_images/EmpiricalProbabilities_11_0.png

It looks like we don’t have a very clean category system here - ‘dress’ is a category, but so are additional categories of dresses and skirts.

Basic / Marginal Probabilities

Let’s consider garment fit (the fit column). This takes on one of three values (the elementary events):

  • fit

  • small

  • large

Let’s define each of these as an event, a singleton set:

  • \(A_{\mathrm{f}} = \{\mathrm{fit}\}\)

  • \(A_{\mathrm{s}} = \{\mathrm{small}\}\)

  • \(A_{\mathrm{l}} = \{\mathrm{large}\}\)

The set “not fit” is \(A_{\mathrm{f}}^c = \{\mathrm{small}, \mathrm{large}\}\).

Let’s compute an empirical value \(\P[A_{\mathrm{f}}] = \P[\mathrm{fit}]\) from the data. This is a fraction:

\[\frac{\text{# of rentals that fit}}{\text{# of rentals}}\]

The basic principle is in two steps:

  • create a logical series that has a non-null element for each element of the denominator, and is True for elements in the numerator

  • compute the fraction of these elements that are True - pandas.Series.mean() and numpy.mean() are both very good for this

Let’s do that now:

fits = rtr['fit'] == 'fit'
fits.mean()
0.7377949975070633

73.8% of garments fit — \(\P[A_{\mathrm{f}}] \approx 0.738\).

We can compute the complementary probability \(\P[A_{\mathrm{f}}^c]\):

not_fits = rtr['fit'] != 'fit'
not_fits.mean()
0.26220500249293666

You can confirm that \(\P[A_{\mathrm{f}}^c] = 1 - \P[A_{\mathrm{f}}]\).

We call \(\P[A_{\mathrm{f}}]\) a marginal probability, even though we did not use marginalization to compute it. The probability is what it is; its status as a marginal probability is based on the definition that it is the probability that a garment fits, without conditioning on any other criteria. Marginalization is a way to compute a marginal probability from conditional probabilities, but there are other ways as well, as we have seen.

There are a few ways to interpret about this probability:

  • Probability is just a mathematical construct, and computing the fraction of garments that fit is compatible with that usage.

  • If we randomly pick a garment rental from the data set, the probability that you pick a rental that fit is \(\P[A_{\mathrm{f}}^c]\).

  • If our sample is representative, and we assume a uniform prior (no prior knowledge about the probability a garment fitting), it is an estimate of the probability that a future rental will fit.

Joint and Conditional Probabilities

Let’s now look at conditional probabilities. Suppose we want to compute \(\P[A_{\mathrm{f}} | C_{\mathrm{g}}]\): the probability that a garment fits, given that it is a gown. This probability can be expanded as:

\[\P[A_{\mathrm{f}} | C_{\mathrm{g}}] = \frac{\P[A_{\mathrm{f}}, C_{\mathrm{g}}]}{\P[C_{\mathrm{g}}]} = \frac{\text{# of gowns that fit}}{\text{# of gowns}}\]

To do this, we need to compute the # of gowns, and the # of gowns that fit. There are at least two different ways to do this; one is to compute joint and marginal probabilities separately. We can use the Python & operator to compute the conjunction of two series, and the joint probability \(\P[A_{\mathrm{f}}, C_{\mathrm{g}}]\):

is_gown = rtr['category'] == 'gown'
is_fitting_gown = is_gown & fits
is_fitting_gown.mean()
0.18269590327405683

This is the probability that a rental is both a gown, and it fits. The complement contains all rentals that are for garments not labeled as gowns, and all gowns that do not fit.

We can then divide by the probability of a gown (\(\P[C_{\mathrm{g}}]\), computed by is_gown.mean()):

is_fitting_gown.mean() / is_gown.mean()
0.7926139564227935

We can see that the probability of a garment fitting, given that it is a gown, is somewhat higher than the marginal probability of a garment fitting. We don’t know if this is statistically significant or not though.

The other way we can do this computation is to limit our data set to only gowns, and then compute the probability of a garment fitting within that subset. The previous computation calculated the conditional probability by computing the joint and marginals, and using the relationship betwen joint and conditional probability to compute the conditional; this version of the calculation will compute the conditional probability \(\P[A_{\mathrm{f}} | C_{\mathrm{g}}]\) directly.

Since we already have a boolean series that is True when the garment is a gown, we can use that to subset:

gowns = rtr[is_gown]
gowns
fit user_id bust size item_id weight rating rented for review_text body type review_summary category height size age review_date
1 fit 273551 34b 153475 132lbs 10.0 other I rented this dress for a photo shoot. The the... straight & narrow I felt so glamourous!!! gown 5' 6" 12 36.0 June 18, 2013
4 fit 151944 34b 616682 145lbs 10.0 wedding I have always been petite in my upper body and... athletic Was in love with this dress !!! gown 5' 9" 12 27.0 September 26, 2016
8 fit 166228 36d 1729232 NaN 10.0 formal affair I was nervous of it looking cheap when it arri... full bust Great for black tie event! gown 5' 6" 21 27.0 June 27, 2016
9 fit 154309 32b 1729232 114lbs 10.0 formal affair The dress was very flattering and fit perfectl... petite This dress was everything! It was perfect for ... gown 5' 3" 1 33.0 October 17, 2016
14 fit 721308 34b 123793 118lbs 10.0 formal affair Fit great, super flattering athletic Stunning gown. Wore this for heart ball and re... gown 5' 5" 2 32.0 May 29, 2014
... ... ... ... ... ... ... ... ... ... ... ... ... ... ... ...
192506 fit 78599 NaN 672419 136lbs 10.0 party This was one of my backup dresses, and it work... hourglass The customer service impeccable! gown 5' 3" 12 35.0 April 12, 2017
192513 small 389288 34b 765872 145lbs 10.0 wedding I ordered a 4L and it was perfect length but w... athletic Besides this dress just being stunning and sex... gown 5' 4" 17 34.0 June 30, 2015
192515 fit 378332 34d 125465 NaN 10.0 wedding This was my backup dress to a black tie weddin... NaN So Flowy and Fun! gown 5' 7" 13 31.0 June 10, 2016
192524 fit 62024 32d 152662 130lbs 6.0 date This was my backup dress I didn't end up weari... hourglass Pretty but kind of bland gown 5' 9" 4 35.0 April 23, 2014
192543 fit 123612 36b 127865 155lbs 10.0 wedding This dress was wonderful! I had originally pla... athletic I wore this to a beautiful black tie optional ... gown 5' 6" 16 30.0 August 29, 2017

44381 rows × 15 columns

Much smaller data set. What’s the probability of fitting (\(\P[A_{\mathrm{f}} | C_{\mathrm{g}}]\))?

np.mean(gowns['fit'] == 'fit')
0.7926139564227935

Here I used numpy.mean() instead of the Pandas mean, just to demonstrate using it, and because its parentheses are slightly friendlier.

We can see these are the same probability. Which method you want to use depends on the data you are working with and the other computations you’re doing — use what is natural and makes sense in the context of your analysis.

Conclusion

This notebook has demonstrated some basic computations for empirically estimating probabilities. You can do quite a bit more, by:

  • subsetting data

  • using the & (and / conjunction / intersection) and | (or / disjunction / union) operators to combine boolean series

See the Tricks with Boolean Series notebook for more about doing things with Boolean series.