The data analytics and science community has long been dominated by the R programming language. While quite capable, R targets a niche audience. It begins to lose its appeal in the broader field of software development. In other words, no data project is an island. The modern web has made it almost mandatory that data be accessible anywhere, at any time and by anyone. The solution is to implement data solutions using a more general-purpose language.
Python has become a popular alternative to R for implementing data solutions. Python has a very small footprint with a simple, clean syntax that is easy to use and remember. Python has a very active open-source community that is contributing packages for a wide variety of tasks, including data science and analytics. And finally, Python is cross-platform. It works equally well on the three major platforms: macOS, Linux, and even Windows.
This guide is not for Python novices. It already assumes basic Python experience. However, it should be noted that someone with significant experience in another language (i.e., Java, C#) should be able to pick up Python quickly. The concepts are largely the same. Syntax is the only major difference.
The term ‘PyData’ is not official but rather a colloquial title that refers to a collection of commonly used Python packages and tools within the data science community. So while it means different things to different Pythonistas, there are a few generally accepted members of this collection. This guide will specifically focus on those and then mention several others at the end.
numpy package handles numerical computing tasks using the Python language. Now, don't let the phrase numerical computing frighten you. You'll soon see how
numpy handles most of the hard work for you. That hard work consists largely of manipulating multidimensional data structures.
Consider the task of matrix multiplication. Multiplying a 2x3 matrix with a 3x4 matrix will, after 24 multiplication operations and 8 addition operations, yield a 2x4 matrix. And this is easily done with a few lines of Python. But recall that Python is an interpreted language and can be quite slow. While it will be tough to observe a performance hit from the above example, higher dimensional data structures will quickly overwhelm pure Python.
This is why
numpy is largely implemented in C. The C code compiles to native, platform specific binaries that run close to the metal, eliminating the performance issues with pure Python. But
numpy still provides Python bindings to the API implemented in C. This gives the developer the best of both worlds: the performance of C and the ease of use of Python.
The central data structure of
numpy is the
ndarray. The “nd” stands for “n-dimensional”, so the
ndarray is capable of representing multidimensional data structures. Creating an
ndarray from a Python list is simple:
1import numpy as np 2data = list(range(20)) 3arr = np.array(data)
np alias for
numpy is a common convention. You'll see it used in a lot of PyData applications. Displaying the value of
arr will yield:
1array([ 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19])
This looks a lot like a Python list. And the
ndarray can be used as a substitute for a list much of the time. But it has learned a few new tricks. For example, the
1mat = arr.reshape(5, 4)
This will return an
ndarray with the dimensions passed to it, in this case 5 rows and 4 columns:
1array([[ 0, 1, 2, 3], 2 [ 4, 5, 6, 7], 3 [ 8, 9, 10, 11], 4 [12, 13, 14, 15], 5 [16, 17, 18, 19]])
While this looks like a nested list in Python, it's actually an
ndarray. Indexing an
ndarray has also learned some new tricks. For example, to access the value in row 1 and column 2, you would write this code with Python lists:
This works with an
ndarray as well, but there is a shortcut that reduces some of the syntactic noise.
The cool thing about
ndarray indexing is that you can select every value in a dimension with a colon, like so:
This effectively selects column two in
mat. The colon says to select every row, and then index two in each row.
Another benefit of
numpy is the idea of broadcasting. Consider the obligatory example of squaring a list of integers:
1my_list = list(range(10)) 2 3squares =  4for value in my_list: 5 sqaure_value = value ** 2 6 squares.append(square_value) 7# or 8squares = [value ** 2 for value in my_list]
Even though this example can be reduced to a single line of code with the list comprehension, it requires some thought to translate the Python code to the mathematical expression it represents. Using an
ndarray to do the same thing is much more readable:
1my_array = np.array(list(range(10)) 2 3square_array = my_array ** 2
numpy scales broadcasting to high dimensions as well:
1my_array = np.random.randint(0, 10, size=(3, 2)) 2square_array = my_array ** 2
random module in
numpy mirrors the
random module in the Python standard library with some extras to accommodate the
ndarray. In the above example,
randint will return a 3x2
ndarray with integer values from 0 to 9.
Operations on more than one
ndarray are just as easy:
1A = np.random.randint(0, 10, size=(3, 2)) 2B = np.random.randint(0, 10, size=(3, 2)) 3mat_sum = A + B
And that matrix multiplication issue mentioned at the beginning of this section?
1A = np.random.randint(0, 10, size=(2, 3)) 2B = np.random.randint(0, 10, size=(3, 4)) 3dot_prod = A.dot(B)
ndarray can be filtered using simple Boolean expressions:
1arr = np.array(list(range(20))) 2odds = arr[arr % 2 == 1]
1array([ 1, 3, 5, 7, 9, 11, 13, 15, 17, 19])
This just scratches the surface of
numpy. There are many other useful features (check out indexing an
ndarray) that this guide cannot cover.
Hopefully you can see how powerful
numpy is and how much time it can save. But at the end of the day, the
ndarray is nothing more than a big glob of numbers. This is because there is no metadata associated with an
ndarray. To work with data in a more intellectually palatable context, we turn to
pandas API is what you are more likely to spend time with. While
numpy handles the core operations on numeric values,
pandas mimics large parts of R. This again gives you the best of both worlds: the power of R with the syntax of Python.
The central data structure in
pandas is the
DataFrame. It mimics the R data structure of the same name. In
DataFrame is a two-dimensional data structure with named columns and indexed rows. In other words, you can conceptually think of it as a table. The values of a
DataFrame are represented by an
ndarray, and a
DataFrame can be created from an
1import numpy as np 2import pandas pd 3 4raw = np.random.randint(0, 10, size=(5, 3)) 5df = pd.DataFrame(data=raw)
The result is
1 0 1 2 20 3 4 0 31 1 8 6 42 0 5 4 53 8 5 8 64 0 1 2
The top row of headers and the left most column of indexes were generated by
pandas. The headers can be modified by setting the
1df.columns = ['Foo', 'Bar', 'Baz']
1 Foo Bar Baz 20 3 4 0 31 1 8 6 42 0 5 4 53 8 5 8 64 0 1 2
The columns can also be accessed by name:
10 3 21 1 32 0 43 8 54 0 6Name: Red, dtype: int64
While bearing a strong resemblance to a
DataFrame, this is actually a
Series is an
ndarray with an index. The column names of the
DataFrame can be used as fields as well:
DataFrame can also be filtered.
1df[df.Foo == 0]
1 Foo Bar Baz 22 0 5 4 34 0 1 2
Notice that a
DataFrame is returned containing the rows in which the value of
Foo is zero.
DataFrame supports dynamically adding a new column:
1df['New'] = np.random.randint(0, 10, size=5)
1 Foo Bar Baz New 20 3 4 0 0 31 1 8 6 9 42 0 5 4 4 53 8 5 8 6 64 0 1 2 2
DataFrame exposes a number of statistical methods such as
1Foo 12 2Bar 23 3Baz 20 4New 21 5dtype: int64
The result is a
Series with the values of the sums of the columns, and the index is the column names. To sum along the rows, add the
axis= keyword argument with a value of 1 (the default is 0).
1df['Total'] = df.sum(axis=1)
1 Foo Bar Baz New Total 20 3 4 0 0 7 31 1 8 6 9 24 42 0 5 4 4 13 53 8 5 8 6 27 64 0 1 2 2 5
You can also change the index:
1df.index = [5, 10, 15, 20, 25]
1 Foo Bar Baz New Total 25 3 4 0 0 7 310 1 8 6 9 24 415 0 5 4 4 13 520 8 5 8 6 27 625 0 1 2 2 5
Accessing the rows of the
DataFrame by index can be done using
1Foo 0 2Bar 5 3Baz 4 4New 4 5Total 13 6Name: 15, dtype: int64
And by zero-based position with
Again, this is just scratching the surface of the capabilities of
It is obviously easier to use
pandas over the
ndarray. But most of the time,
DataFrames will be very large. Presenting a
DataFrame to a knowledge worker or decision maker who needs information at a glance is not an option. For this, we turn to the visualization, which is a graphical representation of data and information. Think of it as another level of abstraction.
In the PyData stack, the
matplotlib package handles visualizations using Python. The
matplotlib API is big, in fact it is huge. I can't begin to cover it in this guide. In fact, I have a two-hour course on Pluralsight about
matplotlib that only covers an introduction to the package. But here are the basics.
First, the module where much of the API resides is
matplotlib.pyplot. This is often shortened to
plt with an alias. Inside of the
plt module are different functions to create different types of visualizations, such as
plot to create a line chart:
1import numpy as np 2import matplotlib.pyplot as plt 3 4data = np.random.randint(0, 10, 10) 5plt.plot(data)
Creating a bar chart is simple as well:
Here's a pie chart:
These visualizations all render the same data in different charts. With the exception of the bar chart, the functions take an
ndarray or any Python iterable. For the
bar function, the first argument is the positions of the bars. To space the bars evenly, just use a list of sequential integers.
And here is a fancy scatter chart.
1x = np.random.randn(25) 2y = np.random.randn(25) 3colors = np.random.uniform(size=25) 4sizes = np.random.uniform(size=25) * 1000 5plt.scatter(x, y, c=colors, s=sizes, cmap='Blues')
The scatter chart plots point in two dimensions. The keyword arguments
s= represent the color and size of the points to show the intensity of the values. The
cmap= keyword argument represents the available colors to use. There are many color maps predefined in
You could change the values of the axes in the charts:
1import string 2plt.plot(data) 3labels = list(string.ascii_uppercase[:len(data)]) 4plt.xticks(list(range(len(labels))), labels)
Similar to the
bar function, the
xticks function expects the first argument to be the locations of the ticks and the second to be the new labels.
The are a number of keyword arguments in
matplotlib to get the appearance of visualizations just the way you want them. You already saw some of these for the scatter chart. The line chart also provides the following:
1plt.plot(data, c='red', ls='--')
The documentation covers the complete set of keyword arguments for each visualization. As you might have guessed, there are too many to cover here.
The next one is a little different. It doesn't expose an API, but instead is a tool. The process of writing code in data science is not like that of most software development. In data science, you set up an experiment, set some initial parameters, run the experiment, collect the results, tweak the parameters, and repeat. This process thrives when you can get rapid feedback, and it also requires different tools.
Jupyter Notebook takes the concept of the Python REPL and goes wild with it. Instead of a platform-specific command line interface, Jupyter Notebook is like a Python session running on a web server. And you use a web browser to access that Python session. This has several advantages. You don't have to be running a specific operating system or even device to access Jupyter Notebook. You don't even need Python installed!
Jupyter Notebook organizes code into cells. A cell may be executed by pressing Shift-Enter.
Jupyter Notebook also persists everything to a file in JSON. This way you can close a notebook and reopen it in the state where you last left it. It also makes sharing experiments much easier. With the Python REPL, when you close it, all of your work is lost.
I've prepared a sample notebook illustrating the concepts discussed in the previous three sections running on a service from Google called Google Colab.
Jupyter Notebook can also be installed locally via
pip, but Colab is a great way to get started. And it's not a toy either. You can actually get access for free to GPUs and run jobs for up to 12 hours.
If you'd like to know more about Jupyter Notebook, check out my video course on Pluralsight: Getting Started with Jupyter Notebook and Python.
There are a lot of other packages often used with the packages discussed, but maybe not as often. Here are a few:
sklearn): This is a machine learning API. It implements many common use cases for machine learning, such as classification and regression. Generally, it is a great choice for someone who does not need to get into the details of machine learning. You can read a Pluralsight Guide I authored about Scikit-learn for more information about it.
sklearn implements common tasks in machine learning,
statsmodels does the same for statistics. It provides an easy-to-use API that avoids the details of implementation.
Sympy: Data science will often require complex mathematical expression. While these can be written using Python, it can be difficult to read them. The
sympy package is a computer algebra system that makes writing mathematical expressions in Python simpler.
NLTK: This is a package for implementing tasks in natural language processing. It includes an API for tagging parts of speech, parsing and tokenizing text and over 50 popular corpora (bodies of text). The
nltk package even has a free, online, full-length ebook that covers the package in detail.
Data science and analytics include a diverse range of topics and skills. While R has been a popular choice of data scientists for years, in the larger field of software engineering, developers want a more general-purpose language. By using Python, developers have access to a powerful, yet simple language that avoids the quirks of R. Thanks to the open-source Python community, developers can choose from a wide selection of packages for implementing solutions to data science problems.