We will start today with the interactive environment that we will be using often through the course: the Jupyter Notebook.
We will walk through the following steps together:
Download miniconda (be sure to get Version 3.6) and install it on your system (hopefully you have done this before coming to class)
Use the conda command-line tool to update your package listing and install the IPython notebook:
Update conda's listing of packages for your system:
$ conda update condaInstall IPython notebook and all its requirements
$ conda install jupyter notebookNavigate to the directory containing the course material. For example:
$ cd ~/courses/CSE583/You should see a number of files in the directory, including these:
$ ls
...
Breakout-Simple-Math.ipynb
Lecture-Python-And-Data.ipynb
...Type jupyter notebook in the terminal to start the notebook
$ jupyter notebookIf everything has worked correctly, it should automatically launch your default browser
Click on Lecture-Python-And-Data.ipynb to open the notebook containing the content for this lecture.
With that, you're set up to use the Jupyter notebook!
Now that we have the Jupyter notebook up and running, we're going to do a short breakout exploring some of the mathematical functionality that Python offers.
Please open Breakout-Simple-Math.ipynb, find a partner, and make your way through that notebook, typing and executing code along the way.
In addition to Python's built-in modules like the math module we explored above, there are also many often-used third-party modules that are core tools for doing data science with Python.
Some of the most important ones are:
numpy: Numerical PythonNumpy is short for "Numerical Python", and contains tools for efficient manipulation of arrays of data. If you have used other computational tools like IDL or MatLab, Numpy should feel very familiar.
scipy: Scientific PythonScipy is short for "Scientific Python", and contains a wide range of functionality for accomplishing common scientific tasks, such as optimization/minimization, numerical integration, interpolation, and much more. We will not look closely at Scipy today, but we will use its functionality later in the course.
pandas: Labeled Data Manipulation in PythonPandas is short for "Panel Data", and contains tools for doing more advanced manipulation of labeled data in Python, in particular with a columnar data structure called a Data Frame. If you've used the R statistical language (and in particular the so-called "Hadley Stack"), much of the functionality in Pandas should feel very familiar.
matplotlib: Visualization in PythonMatplotlib started out as a Matlab plotting clone in Python, and has grown from there in the 15 years since its creation. It is the most popular data visualization tool currently in the Python data world (though other recent packages are starting to encroach on its monopoly).
Because the above packages are not included in Python itself, you need to install them separately. While it is possible to install these from source (compiling the C and/or Fortran code that does the heavy lifting under the hood) it is much easier to use a package manager like conda. All it takes is to run
$ conda install numpy scipy pandas matplotliband (so long as your conda setup is working) the packages will be downloaded and installed on your system.
In [1]:
!ls
uncomment this to download the data:
In [2]:
# !curl -o pronto.csv https://data.seattle.gov/api/views/tw7j-dfaw/rows.csv?accessType=DOWNLOAD
In [3]:
import pandas
Because we'll use it so much, we often import under a shortened name using the import ... as ... pattern:
In [4]:
import pandas as pd
Now we can use the read_csv command to read the comma-separated-value data:
In [5]:
data = pd.read_csv('pronto.csv')
Note: strings in Python can be defined either with double quotes or single quotes
The head() and tail() methods show us the first and last rows of the data
In [6]:
data.head()
Out[6]:
In [7]:
data.tail()
Out[7]:
The shape attribute shows us the number of elements:
In [8]:
data.shape
Out[8]:
The columns attribute gives us the column names
In [9]:
data.columns
Out[9]:
The index attribute gives us the index names
In [10]:
data.index
Out[10]:
The dtypes attribute gives the data types of each column:
In [11]:
data.dtypes
Out[11]:
Access columns by name using square-bracket indexing:
In [12]:
data.columns
Out[12]:
In [13]:
data['tripduration']
Out[13]:
Mathematical operations on columns happen element-wise:
In [14]:
data['tripduration'] / 60
Out[14]:
Columns can be created (or overwritten) with the assignment operator. Let's create a tripminutes column with the number of minutes for each trip
In [15]:
data['tripminutes'] = data['tripduration'] / 60
In [16]:
data.head()
Out[16]:
Note that this manipulation only modifies the data frame in memory; if you want to save the modified dataframe to CSV, you can use the to_csv() method:
In [17]:
data.to_csv('pronto-new.csv')
In [18]:
!ls
More complicated mathematical operations can be done with tools in the numpy package:
In [19]:
import numpy as np
np.exp(data['tripminutes'])
Out[19]:
One trick to know when working with columns of times is that Pandas DateTimeIndex provides a nice interface for working with columns of times.
For a dataset of this size, using pd.to_datetime and specifying the date format can make things much faster (from the strftime reference, we see that the pronto data has format "%m/%d/%Y %I:%M:%S %p"
In [20]:
data['starttime'].head()
Out[20]:
In [21]:
pd.to_datetime(data['starttime'].head())
Out[21]:
In [22]:
times = pd.DatetimeIndex(data['starttime'].head())
times.dayofweek
Out[22]:
In [23]:
data['starttime']
Out[23]:
In [24]:
times = pd.DatetimeIndex(pd.to_datetime(data['starttime'], format="%m/%d/%Y %I:%M:%S %p"))
(Note: you can also use infer_datetime_format=True in most cases to automatically infer the correct format, though due to a bug it doesn't work when AM/PM are present)
With it, we can extract, the hour of the day, the day of the week, the month, and a wide range of other views of the time:
In [25]:
times.dayofweek
Out[25]:
In [26]:
times.month
Out[26]:
In [27]:
times
Out[27]:
In [28]:
data.head()
Out[28]:
Pandas includes an array of useful functionality for manipulating and analyzing tabular data. We'll take a look at two of these here.
The pandas.value_counts returns statistics on the unique values within each column.
We can use it, for example, to break down rides by gender:
In [29]:
pd.value_counts(data['gender'])
Out[29]:
Or to break down rides by age:
In [30]:
pd.value_counts(data['birthyear'])
Out[30]:
By default, the values rather than the index are sorted. Use sort=False to turn this behavior off:
In [31]:
pd.value_counts(data['birthyear'], sort=False)
Out[31]:
We can explore other things as well: day of week, hour of day, etc.
In [32]:
pd.value_counts(times.dayofweek, sort=False)
Out[32]:
In [33]:
pd.value_counts(times.month, sort=False)
Out[33]:
In [34]:
pd.value_counts(data['gender'], dropna=False)
Out[34]:
One of the killer features of the Pandas dataframe is the ability to do group-by operations. You can visualize the group-by like this (image borrowed from the Python Data Science Handbook)
In [35]:
from IPython.display import Image
Image('split_apply_combine.png')
Out[35]:
The simplest version of a groupby looks like this, and you can use almost any aggregation function you wish (mean, median, sum, minimum, maximum, standard deviation, count, etc.)
<data object>.groupby(<grouping values>).<aggregate>()for example, we can group by gender and find the average of all numerical columns:
In [36]:
data.groupby('gender').count()
Out[36]:
In [37]:
data.groupby('gender').mean()
Out[37]:
It's also possible to indes the grouped object like it is a dataframe:
In [38]:
data.groupby('gender')['tripminutes'].mean()
Out[38]:
You can even group by multiple values: for example we can look at the trip duration by time of day and by gender:
In [39]:
data.groupby([times.hour, 'gender'])['tripminutes'].mean()
Out[39]:
The unstack() operation can help make sense of this type of multiply-grouped data. What this technically does is split a multiple-valued index into an index plus columns:
In [40]:
grouped = data.groupby([times.hour, 'gender'])['tripminutes'].mean().unstack()
grouped
Out[40]:
pandasOf course, looking at tables of data is not very intuitive.
Fortunately Pandas has many useful plotting functions built-in, all of which make use of the matplotlib library to generate plots.
Whenever you do plotting in the IPython notebook, you will want to first run this magic command which configures the notebook to work well with plots:
In [41]:
%matplotlib inline
Now we can simply call the plot() method of any series or dataframe to get a reasonable view of the data:
In [42]:
grouped.plot()
Out[42]:
In [43]:
import matplotlib.pyplot as plt
plt.style.use('seaborn')
In [44]:
grouped.plot()
Out[44]:
For example, we can create a histogram of trip durations:
In [45]:
grouped.plot.bar()
Out[45]:
If you'd like to adjust the x and y limits of the plot, you can use the set_xlim() and set_ylim() method of the resulting object:
In [46]:
ax = grouped.plot.bar()
ax.set_xlim(-1, 10)
Out[46]:
In [47]:
data['month'] = times.month
In [48]:
ax = data.groupby('month')['trip_id'].count().plot.bar();
Split this plot by gender. Do you see any seasonal ridership patterns by gender?
In [49]:
data.groupby(['month','gender'])['trip_id'].count().unstack().plot();
Split this plot by user type. Do you see any seasonal ridership patterns by usertype?
In [50]:
data.groupby(['month','usertype'])['trip_id'].count().unstack().plot();
Repeat the above three steps, counting the number of rides by time of day rather thatn by month.
In [51]:
data['hour'] = times.month
In [52]:
ax = data.groupby('hour')['trip_id'].count().plot.bar();
In [53]:
data.groupby(['hour','gender'])['trip_id'].count().unstack().plot();
In [54]:
data.groupby(['hour','usertype'])['trip_id'].count().unstack().plot();
Are there any other interesting insights you can discover in the data using these tools?
A writeup of the analysis that Jake did on the first year of this data, using these same tools, can be found here: http://nbviewer.jupyter.org/github/jakevdp/ProntoData/blob/master/ProntoData.ipynb
In the homework this week, you will have a chance to apply some of these patterns to a brand new (but closely related) dataset.