Manipulating and Analyzing Data With Pandas

Manipulating and analyzing data with pandas
Céline Comte
Nokia Bell Labs France & Télécom ParisTech
Python Academy - May 20, 2019

Introduction
• Pandas: Python Data Analysis Library
• “An open source, BSD-licensed library providing high-performance,

easy-to-use data structures and data analysis tools for the Python
programming language”
(https://pandas.pydata.org/)
• Sponsored by NumFOCUS, a non-profit organization in the US

(like NumPy, Matplotlib, Jupyter, and Julia)
• Used in StatsModel, sklearn-pandas, Plotly, IPython, Jupyter, Spyder

(http://pandas-docs.github.io/pandas-docs-travis/ecosystem.html)
2/50 © 2019 Nokia Public

Side remark: BSD licenses
• BSD = Berkeley Software Distribution
The first software (an OS actually) to be distributed under BSD license
“Permissive” license → can be used in a proprietary software
(https://github.com/pandas-dev/pandas/blob/master/LICENSE)
Introduction
• Built on top of NumPy
• Part of the SciPy ecosystem

(Scientific Computing Tools for Python)
• Version history (https://pandas.pydata.org/
community.html#history-of-development)
− Project initiated in 2008
− Oldest version in the doc:
0.4.1 (September 2011)
− Current version: 0.24.2 (March 2019)

Objectives of the presentation
• Explain when one can benefit from using pandas
• Describe the data structures in pandas

Series 1-dimensional array with labels
DataFrame 2-dimensional array with labels
Panel 3-dimensional array with labels
(deprecated since version 0.20.0)
• Review the data analysis tools in pandas
− Import and export data
− Select data and reshape arrays
− Merge, join, and concatenate arrays
− Visualize data
− …

Two distinct questions
• What is the advantage as a programmer?

Addressed in this presentation.
• What is the speed of the obtained code?

Not addressed in this presentation. Two brief comments:
− Pandas is an overlay on top of NumPy.
Because of this, it may have a performance cost.
− “pandas is fast. Many of the low-level algorithmic bits have been
extensively tweaked in Cython code. However, as with anything
else generalization usually sacrifices performance.”
(http://pandas.pydata.org/pandas-docs/stable/getting_started/overview.html)

Outline
NumPy
Data structures in pandas

Series
DataFrame
Data analysis tools in pandas (10 minutes to pandas)

(http://pandas.pydata.org/pandas-docs/stable/getting_started/10min.html)

Outline
NumPy

Series
DataFrame


Overview
• NumPy: Numeric Python or Numerical Python

(https://www.datacamp.com/community/tutorials/python-numpy-tutorial)
• Data structure: A fixed-size multidimensional array object called “ndarray”,

for “N-dimensional array”, or just “array”
• Tools: “Routines for fast operations on arrays, including mathematical, logical,

shape manipulation, sorting, selecting, I/O, discrete Fourier transforms, basic
linear algebra, basic statistical operations, random simulation and much
more.” (https://www.numpy.org/devdocs/user/whatisnumpy.html)
• Two important notions: vectorizing and broadcasting

Numpy ndarray structure
(https://www.numpy.org/devdocs/user/whatisnumpy.html)
(https://www.numpy.org/devdocs/reference/arrays.ndarray.html)
• A fixed-size multidimensional array object
• “NumPy arrays have a fixed size at creation, unlike Python lists (which can
grow dynamically). Changing the size of an ndarray will create a new array and
delete the original.”
• ”The elements in a NumPy array are all required to be of the same data type,
and thus will be the same size in memory. The exception: one can have arrays
of (Python, including NumPy) objects, thereby allowing for arrays of different
sized elements.“
• Advantage of this rigidity: (usually) contiguous block of memory
→ Faster code

Create an ndarray
# define an array with two axes
> a = np.array([[3., 0.], [20., 230.], [21., 275.]])
array([[ 3., 0.],
[ 20., 230.],
[ 21., 275.]])
Axis 1
> a[2,0] # same result as a[2][0]
0 1
21.0
Axis 0
0 3. 0.
1 20. 230.
> a[:2, :] # returns a view of the array
array([[ 3., 0.], 2 21. 275.
[ 20., 230.]])

Some attributes
> a.shape
(3, 2)
# axis 0 is of length 3, axis 1 is of length 2
> a.dtype
dtype('float64') Axis 1
# data-type, specifies how to interpret each item 0 1
# (inferred from data if unspecified)
Axis 0
0 3. 0.
1 20. 230.
> a.itemize 2 21. 275.
8 # the size of each element of the array,
# in bytes (8 x 8 = 64)

View: different numpy object but same data
# create a view of the array

> b = a[:2, :]
array([[ 3., 0.],
[ 20., 230.]])
> b[0,0] = 0. Axis 1

0 1
> a
Axis 0
0 3. 0.
array([[ 0., 0.], 1 20. 230.
[ 20., 230.], 2 21. 275.
[ 21., 275.]])

Reshape an array
(https://www.numpy.org/devdocs/user/quickstart.html#changing-the-shape-of-an-array)
The reshape method returns its argument with a modified shape, whereas the
resize method modifies the array itself:
> a.resize(2,3)
> b = a.reshape(2,3)
> a
array([[ 3., 0., 20.],
array([[ 3., 0., 20.],
[230., 21., 275.]])
[230., 21., 275.]])
> a
array([[ 3., 0.],
> np.resize(a, (2,3))
[ 20., 230.],
array([[ 3., 0., 20.],
[ 21., 275.]])
[230., 21., 275.]])

Functionalities
• “Routines for fast operations on arrays, including mathematical, logical, shape
manipulation, sorting, selecting, I/O, discrete Fourier transforms, basic linear
algebra, basic statistical operations, random simulation and much more.”
(https://www.numpy.org/devdocs/user/whatisnumpy.html)
• Summing the elements of an array

> a.sum() Axis 1
549.0 0 1
> a.sum(axis=0)
Axis 0
0 3. 0.
array([ 44., 505.]) 1 20. 230.
2 21. 275.
• Taking the maximum of an array
> a.max() # and, similarly, a.max(axis=0)
275.0

Structured arrays: addressing columns by name
(https://scipy-cookbook.readthedocs.io/items/Recarray.html)
> a = np.array([(3., 0.), (20., 230.), (21., 275.)],

dtype=np.dtype([('Age', int), ('Weight', float)]))
array([( 3, 0.), (20, 230.), (21, 275.)],
dtype=[('Age', '<i8'), ('Weight', '<f8')])
Axis 1
> a['Age']
array([ 3, 20, 21]) 0 1
Axis 0
0 3 0.
1 20 230.
Here, a is a 1-dimensional array with tuple elements.
2 21 275.
> a[0]
(3, 0.)

Outline
NumPy

Series
DataFrame


Philosophy
• Arrays with smaller dimensions

Variables
− Series: 1-dimensional
Observations
− DataFrames: 2-dimensional Age Weight
Bei Bei 3
• Give a semantical meaning to the axes Mei Xiang 20 230.
− Columns ≃ Variables Tian Tian 21 275.
− Lines ≃ Observations
• Other functionalities:
− Missing data: Identified by NaN (np.nan).
− Mutability: Add and remove columns in a DataFrame
− Data alignment: Combine data based on the indices

Data structures
Series “One-dimensional ndarray with axis labels (including time series).”

(https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.
Series.html)
DataFrame “Two-dimensional size-mutable, potentially heterogeneous tabular

data structure with labeled axes (rows and columns).
Arithmetic operations align on both row and column labels.
Can be thought of as a dict-like container for Series objects.
The primary pandas data structure.”
(https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.
DataFrame.html)

Series
(https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.Series.html)
Weight (pounds)
• Like ndarrays, the length of a Series cannot be
Axis 0
modified after definition. Mei Xiang 230.
Tian Tian 275.
• Index: Can be of any hashable type.
• Automatic data alignment: “Data alignment is intrisic. Age (years)

The link between labels and data will not be broken Bei Bei 3
unless done so explicitly by you.” (https://pandas.pydata. Mei Xiang 20
org/pandas-docs/stable/getting_started/dsintro.html)
Tian Tian 21
• Missing data: Represented as NaN (np.nan, a float!).
“Statistical methods from ndarray have been overridden
to automatically exclude missing data”.

Creating a Series
> s = pd.Series([3, 20, 21],
index=['Bei Bei', 'Mei Xiang', 'Tian Tian'],
name='Age') Age
Bei Bei 3
Bei Bei 3
Mei Xiang 20
Tian Tian 21 Mei Xiang 20
Name: Age, dtype: int64 Tian Tian 21
> s.array # ``a thin (no copy) wrapper around numpy.ndarray''

<PandasArray>
[3, 20, 21]
Length: 3, dtype: int64

Some attributes
> s = pd.Series([3, 20, 21],
name='Age') Age
Bei Bei 3
> s.dtype # default value: inferred from data
dtype('int64') # usually of type numpy.dtype Mei Xiang 20
Tian Tian 21
> s.name # default value: None
'Age'
> s.index # default value: RangeIndex(start=0, stop=6, step=1)

Index(['Bei Bei', 'Mei Xiang', 'Tian Tian'], dtype='object')

Accessing data
> s = pd.Series([3, 20, 21],
name='Age') Age
Bei Bei 3
> s['Mei Xiang'] # same as s[1]
20 Mei Xiang 20
Tian Tian 21
> s['Mei Xiang':'Tian Tian'] # same as s[1:]
Mei Xiang 20
Tian Tian 21
Name: Age, dtype: int64

Creating a View
> t = s['Mei Xiang':'Tian Tian'] # another Series with the same data
Mei Xiang 20
Tian Tian 21 Age
Bei Bei 3
> t['Tian Tian'] = 22 Mei Xiang 20
Tian Tian 21
> s
Bei Bei 3
Mei Xiang 20
Tian Tian 22

Adding two series (with automatic data alignment)
> u = pd.Series([230., 275.],
index=['Mei Xiang', 'Tian Tian'], Weight
name='Weight')
Mei Xiang 230.0 Mei Xiang 230.
Tian Tian 275.0 Tian Tian 275.
Name: Weight, dtype: float64
> s.add(u) # same as s + u, also the default in NumPy Age

Bei Bei NaN Bei Bei 3
Mei Xiang 250.0 Mei Xiang 20
Tian Tian 296.0
Tian Tian 21
dtype: float64

Adding two series (with automatic data alignment)
> u = pd.Series([230., 275.],
index=['Mei Xiang', 'Tian Tian'], Weight
name='Weight')
Mei Xiang 230.0 Mei Xiang 230.
Tian Tian 275.0 Tian Tian 275.
Name: Weight, dtype: float64
> s.add(u, fill_value=0) Age

Bei Bei 3.0 Bei Bei 3
Mei Xiang 250.0 Mei Xiang 20
Tian Tian 296.0
Tian Tian 21
dtype: float64

DataFrame
(https://pandas.pydata.org/pandas-docs/stable/reference/api/pandas.DataFrame.html)
• “Two-dimensional size-mutable, potentially

heterogeneous tabular data structure with Axis 1
labeled axes (rows and columns).” Age Weight
Axis 0
BB 3
• Mutability: Columns can have different
MX 20 230.
dtypes and can be added and removed,
but they have a fixed size. TT 21 275.
• Semantic: Similar to a table in a relational database.

Like in R language (https://en.wikipedia.org/wiki/R_(programming_language))
− Columns ≃ Variables
− Rows ≃ Observations of these variables

A dictionary of Series
Axis 1
• Column labels ≃ keys, columns ≃ values Age Weight
Axis 0
(https://pandas.pydata.org/pandas-docs/stable/ BB 3
reference/api/pandas.DataFrame.html) MX 20 230.
TT 21 275.
• “You can treat a DataFrame semantically like
a dict of like-indexed Series objects. Getting, ↓
setting, and deleting columns works with the
same syntax as the analogous dict operations”. Age Weight
(https://pandas.pydata.org/pandas-docs/stable/
BB 3 MX 230.
getting_started/dsintro.html)
MX 20 TT 275.
• In particular: access by key, del, pop. TT 21

Creating a DataFrame
> df = pd.DataFrame({'Age': [3, 20, 21], 'Weight': [np.nan, 230., 275.]},
index=['Bei Bei', 'Mei Xiang', 'Tian Tian'])
Age Weight
Bei Bei 3 NaN
Mei Xiang 20 230.0 Axis 1
Tian Tian 21 275.0 Age Weight
Axis 0
BB 3
> df.dtypes # returns a Series MX 20 230.
Age int64 TT 21 275.
Weight float64
dtype: object
In general: list ≃ rows, dictionary ≃ columns.

Other attributes
> df.shape
(3, 2) Axis 1
Age Weight
> df.size
Axis 0
BB 3
6 MX 20 230.
TT 21 275.
> df.columns
Index(['Age', 'Weight'], dtype='object')
> df.index
Index(['Bei Bei', 'Mei Xiang', 'Tian Tian'],
dtype='object')

Accessing data
There are many ways of accessing data (loc, iloc, at, iat).
Axis 1
> df['Age'] # a column (Series)
Bei Bei 3 Age Weight
Axis 0
Mei Xiang 20 BB 3
Tian Tian 21 MX 20 230.
Name: Age, dtype: int64 TT 21 275.
> df['Mei Xiang':'Tian Tian'] # a range of rows (DataFrame)

Weight Age
Mei Xiang 230.0 20
Tian Tian 275.0 21
> df['Mei Xiang'] # KeyError

Summing over columns and rows
Returns a Series
Excludes NaN values by default (skipna=True) Axis 1
> df.sum() # same as df.sum(axis=0) Age Weight
Axis 0
Age 44.0 BB 3
Weight 505.0 MX 20 230.
dtype: float64 TT 21 275.
> df.sum(axis=1)
Bei Bei 3.0
Mei Xiang 250.0
Tian Tian 296.0
dtype: float64

Outline
NumPy

Series
DataFrame


Overview
• Review the data analysis tools provided by pandas.
• The organization and most of the examples of this section come from
the official tutorial 10 minutes to pandas.
• Some examples originate from the User Guide.

(http://pandas.pydata.org/pandas-docs/stable/user_guide/index.html)

Object creation
> dates = pd.date_range('20130101', periods=5)
DatetimeIndex(['2013-01-01', '2013-01-02', '2013-01-03',
'2013-01-04', '2013-01-05'],
dtype='datetime64[ns]', freq='D')
> df = pd.DataFrame(np.random.randn(5, 4), index=dates,

columns=list('ABCD'))
A B C D
2013-01-01 1.501942 -1.459551 -0.376242 0.410211
2013-01-02 0.803188 -0.651458 1.457657 1.575324
2013-01-03 -0.398711 -0.496614 1.032707 0.343666
2013-01-04 0.101690 0.982808 -1.049312 0.535201
2013-01-05 -0.271261 -0.557231 0.307699 0.626503

Viewing Data
> df.head(3) # default value: 5

A B C D
2013-01-01 1.501942 -1.459551 -0.376242 0.410211
2013-01-02 0.803188 -0.651458 1.457657 1.575324
2013-01-03 -0.398711 -0.496614 1.032707 0.343666
> df.tail(3) # default value: 5

A B C D
2013-01-03 -0.398711 -0.496614 1.032707 0.343666
2013-01-04 0.101690 0.982808 -1.049312 0.535201
2013-01-05 -0.271261 -0.557231 0.307699 0.626503

Viewing Data
> df.index
DatetimeIndex(['2013-01-01', '2013-01-02', '2013-01-03', '2013-01-04',
'2013-01-05'], dtype='datetime64[ns]', freq='D')
> df.columns
Index(['A', 'B', 'C', 'D'], dtype='object')
> df.A # same as df['A']

2013-01-01 0.000000
2013-01-02 -0.718574
...
2013-01-05 0.593794
Freq: D, Name: A, dtype: float64

Viewing Data
to_numpy returns an ndarray that contains the data.
May require a copy if the data is heterogeneous.
> a = df.to_numpy()
array([[-2.35406005, -0.31282731, 0.19482154, 1.14387112],
[ 1.70706975, -0.78209048, 0.06241179, -0.00753477],
[-0.21252435, 0.06799263, 1.03563884, -0.67680038],
[ 0.65801543, -0.39368803, 0.5654252 , -1.32672643],
[-1.30699305, -0.06174394, 0.09464223, -0.97696831]])
> a[0,0] = 0
> df['A'][0]
0.0

Selection
Obtain a view of the data.
By label (loc, at), by index (iloc, iat), or by boolean indexing.
> df[df.A > 0] # show the rows where A is positive

A B C D
2013-01-05 0.593794 -0.191118 0.622146 1.325086
> df = df[df > 0] # replaces non-positive values with NaN

A B C D
2013-01-01 NaN 1.112209 0.277689 1.300440
2013-01-02 NaN NaN 1.728119 NaN
2013-01-03 NaN NaN 0.056013 0.970420
2013-01-04 NaN 0.364966 NaN NaN
2013-01-05 0.593794 NaN 0.622146 1.325086
Missing Data
Remove or replace missing data:
• df.dropna: Deletes columns or rows that contain missing values (NaN).
• df.fillna: Fills the NaN with the provided value.
• df.isna or pd.isna(df): Returns a DataFrame of the same size as df with
boolean values that say if the original value in df is NaN.
> df.fillna(value=0)
A B C D
2013-01-01 0.00000 1.171513 0.000000 0.298407
2013-01-02 0.00000 0.893041 2.136786 0.000000
2013-01-03 0.00000 0.030041 0.131783 0.000000
2013-01-04 0.46075 0.000000 0.000000 0.000000
2013-01-05 0.00000 0.953238 0.778675 1.109996

Getting Data In/Out
The basic functions are df.to_csv and df.read_csv.
> df.to_csv('foo.csv')
> df = pd.read_csv('foo.csv', index_col=0)

A B C D
2013-01-01 NaN 1.171513 NaN 0.298407
2013-01-02 NaN 0.893041 2.136786 NaN
2013-01-03 NaN 0.030041 0.131783 NaN
2013-01-04 0.46075 NaN NaN NaN
2013-01-05 NaN 0.953238 0.778675 1.109996

Hierarchical indexing (MultiIndex)
(http://pandas.pydata.org/pandas-docs/stable/user_guide/advanced.html)
Paves the way for higher dimensional data.
Example with two levels of indices:
> arrays = [np.array(['bar', 'bar', 'foo', 'foo']),

np.array(['one', 'two', 'one', 'two'])]
> df = pd.DataFrame(np.random.randn(4, 3), index=arrays)

0 1 2
bar one -0.783896 -1.033699 0.113092
two 0.376749 -0.617641 1.858707
foo one -0.345071 0.288537 0.251429
two 1.391096 -0.053008 -1.290041

Various operations
> df.mean() # same as df.mean(axis=0)
A 0.593794
B 0.738587
C 0.670992
D 1.198649
dtype: float64
Other examples:
sub Substracts another Series or DataFrame (broadcasting)
apply Applies a function such as np.cumsum
value_counts Counts the number of occurrencies of each value

Gather Series or DataFrames
Concatenate:
concat General-purpose function
Concatenate Series or DataFrames along colums or rows
append Append rows to a DataFrame
Equivalent to concat along axis 0
Join / Merge:
merge Database-like join operations

Grouping
> df = pd.DataFrame({'A': ['foo', 'bar', 'foo', 'bar'],
'B': np.random.randn(4), 'C': np.random.randn(4)})
A B C
0 foo -0.124831 1.020952
1 bar -0.755884 1.420176
2 foo -0.736155 -0.229302
3 bar -0.318638 1.232845
> df.groupby('A').sum() # the elements of A are the indices

B C
A
bar -1.074521 2.653020
foo -0.860986 0.791651

Time Series
> rng = pd.date_range('1/1/2012', periods=4, freq='S')
> ts = pd.Series(np.random.randint(0, 500, len(rng)), index=rng)

2012-01-01 00:00:00 55
2012-01-01 00:00:01 163
2012-01-01 00:00:02 31
2012-01-01 00:00:03 167 Support for time zone representation,
Freq: S, dtype: int64 converting to another time zone,
and converting between time span
> ts.resample('2S').sum() representations.
2012-01-01 00:00:00 218
2012-01-01 00:00:02 198
Freq: 2S, dtype: int64

Categoricals
(http://pandas.pydata.org/pandas-docs/stable/user_guide/categorical.html)
Similar to categorical variables used in statistics.

Practical for saving memory and sorting data.
“Examples are gender, social class, blood type, country affiliation, observation
time or rating via Likert scales.”
> s = pd.Series(["a", "b", "c", "a"], dtype="category")

0 a
1 b
2 c
3 a
dtype: category
Categories (3, object): [a, b, c]

Plotting
Based on the API matplotlib.pyplot.
> ts = pd.Series(np.random.randn(1000),
index=pd.date_range('1/1/2000', periods=1000))
> ts.head(3)
2000-01-01 0.310037
2000-01-02 1.747102
2000-01-03 -2.121889
Freq: D, dtype: float64
> ts = ts.cumsum()
> ts.plot()

Resources
Python Data Analysis Library (pandas)
• Documentation
• GitHub awesome-pandas
Links towards videos, cheat sheets, tutorials, books, and papers
→ Video Pandas from the Inside by Stephen Simmons at PyData 2017
About implementation and performance
→ Cheat sheet Data Wrangling with pandas
Numerical Python (NumPy)

• What makes Numpy Arrays Fast: Memory and Strides by Jessica Yung
• From Python to Numpy by Nicolas P. Rougier

Resources
Bei Bei, Mei Xiang, and Tian Tian

• Smithsonian’s National Zoo’s Panda Cams
• Their age and weight were found on Wikipedia

Manipulating and Analyzing Data With Pandas

Uploaded by

Copyright:

Available Formats

Manipulating and Analyzing Data With Pandas

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Manipulating and Analyzing Data With Pandas

Uploaded by

Copyright:

Available Formats

Manipulating and analyzing data with pandas

Python Academy - May 20, 2019

• Pandas: Python Data Analysis Library

• “An open source, BSD-licensed library providing high-performance,

• Sponsored by NumFOCUS, a non-profit organization in the US

• Used in StatsModel, sklearn-pandas, Plotly, IPython, Jupyter, Spyder

2/50 © 2019 Nokia Public

• Built on top of NumPy

• Part of the SciPy ecosystem

4/50 © 2019 Nokia Public

• Describe the data structures in pandas

5/50 © 2019 Nokia Public

• What is the advantage as a programmer?

• What is the speed of the obtained code?

6/50 © 2019 Nokia Public

Data structures in pandas

Data analysis tools in pandas (10 minutes to pandas)

7/50 © 2019 Nokia Public

Data structures in pandas

Data analysis tools in pandas (10 minutes to pandas)

8/50 © 2019 Nokia Public

• NumPy: Numeric Python or Numerical Python

• Data structure: A fixed-size multidimensional array object called “ndarray”,

• Tools: “Routines for fast operations on arrays, including mathematical, logical,

• Two important notions: vectorizing and broadcasting

9/50 © 2019 Nokia Public

• A fixed-size multidimensional array object

10/50 © 2019 Nokia Public

11/50 © 2019 Nokia Public

12/50 © 2019 Nokia Public

# create a view of the array

> b[0,0] = 0. Axis 1

13/50 © 2019 Nokia Public

14/50 © 2019 Nokia Public

• Summing the elements of an array

15/50 © 2019 Nokia Public

> a = np.array([(3., 0.), (20., 230.), (21., 275.)],

16/50 © 2019 Nokia Public

Data structures in pandas

Data analysis tools in pandas (10 minutes to pandas)

17/50 © 2019 Nokia Public

• Arrays with smaller dimensions

18/50 © 2019 Nokia Public

Series “One-dimensional ndarray with axis labels (including time series).”

DataFrame “Two-dimensional size-mutable, potentially heterogeneous tabular

19/50 © 2019 Nokia Public

• Automatic data alignment: “Data alignment is intrisic. Age (years)

20/50 © 2019 Nokia Public

> s.array # ``a thin (no copy) wrapper around numpy.ndarray''

21/50 © 2019 Nokia Public

> s.index # default value: RangeIndex(start=0, stop=6, step=1)

22/50 © 2019 Nokia Public

23/50 © 2019 Nokia Public

24/50 © 2019 Nokia Public

> s.add(u) # same as s + u, also the default in NumPy Age

25/50 © 2019 Nokia Public

> s.add(u, fill_value=0) Age

26/50 © 2019 Nokia Public

• “Two-dimensional size-mutable, potentially