NumPy, which is Travis Oliphant's creation, is the true analytical workhorse of the Python language. It provides the user with multidimensional arrays, along with a large set of functions to operate a multiplicity of mathematical operations on these arrays. Arrays are blocks of data arranged along multiple dimensions, which implement mathematical vectors and matrices. Arrays are useful not just for storing data, but also for fast matrix operations (vectorization), which are indispensable when you wish to solve ad hoc data science problems.

As a convention largely adopted by the Python community, when importing NumPy, it is suggested that you alias it as np:

import numpy as np

We will be doing this throughout the course of this book.

An original project by Travis Oliphant, Pearu Peterson, and Eric Jones, SciPy completes NumPy's functionalities, offering a larger variety of scientific algorithms for linear algebra, sparse matrices, signal and image processing, optimization, fast Fourier transformation, and much more.

The pandas package deals with everything that NumPy and SciPy cannot do. Thanks to its specific object data structures, DataFrames and Series, pandas allows you to handle complex tables of data of different types (which is something that NumPy's arrays cannot do) and time series. Thanks to Wes McKinney's creation, you will be able to easily and smoothly load data from a variety of sources. You can then slice, dice, handle missing elements, add, rename, aggregate, reshape, and finally visualize this data at your will.

Conventionally, pandas is imported as pd:

import pandas as pd

Started as part of the SciKits (SciPy Toolkits), Scikit-learn is the core of data science operations on Python. It offers all that you may need in terms of data preprocessing, supervised and unsupervised learning, model selection, validation, and error metrics. Expect us to talk at length about this package throughout this book. Scikit-learn started in 2007 as a Google Summer of Code project by David Cournapeau. Since 2013, it has been taken over by the researchers at INRA (French Institute for Research in Computer Science and Automation).

A scientific approach requires the fast experimentation of different hypotheses in a reproducible fashion. IPython was created by Fernando Perez in order to address the need for an interactive Python command shell (which is based on shell, web browser, and the application interface), with graphical integration, customizable commands, rich history (in the JSON format), and computational parallelism for an enhanced performance. IPython is our favored choice throughout this book, and it is used to clearly and effectively illustrate operations with scripts and data and the consequent results.

Originally developed by John Hunter, matplotlib is the library that contains all the building blocks that are required to create quality plots from arrays and to visualize them interactively.

You can find all the MATLAB-like plotting frameworks inside the pylab module.

You can simply import what you need for your visualization purposes with the following command:

import matplotlib.pyplot as plt

Previously part of SciKits, statsmodels was thought to be a complement to SciPy statistical functions. It features generalized linear models, discrete choice models, time series analysis, and a series of descriptive statistics as well as parametric and nonparametric tests.

Beautiful Soup, a creation of Leonard Richardson, is a great tool to scrap out data from HTML and XML files retrieved from the Internet. It works incredibly well, even in the case of tag soups (hence the name), which are collections of malformed, contradictory, and incorrect tags. After choosing your parser (basically, the HTML parser included in Python's standard library works fine), thanks to Beautiful Soup, you can navigate through the objects in the page and extract text, tables, and any other information that you may find useful.

Developed by the Los Alamos National Laboratory, NetworkX is a package specialized in the creation, manipulation, analysis, and graphical representation of real-life network data (it can easily operate with graphs made up of a million nodes and edges). Besides specialized data structures for graphs and fine visualization methods (2D and 3D), it provides the user with many standard graph measures and algorithms, such as the shortest path, centrality, components, communities, clustering, and PageRank. We will frequently use this package in Chapter 5, Social Network Analysis.

Conventionally, NetworkX is imported as nx:

import networkx as nx

The Natural Language Toolkit (NLTK) provides access to corpora and lexical resources and to a complete suit of functions for statistical Natural Language Processing (NLP), ranging from tokenizers to part-of-speech taggers and from tree models to named-entity recognition. Initially, the package was created by Steven Bird and Edward Loper as an NLP teaching infrastructure for CIS-530 at the University of Pennsylvania. It is a fantastic tool that you can use to prototype and build NLP systems.

Gensim, programmed by Radim Řehůřek, is an open source package that is suitable for the analysis of large textual collections with the help of parallel distributable online algorithms. Among advanced functionalities, it implements Latent Semantic Analysis (LSA), topic modeling by Latent Dirichlet Allocation (LDA), and Google's word2vec, a powerful algorithm that transforms text into vector features that can be used in supervised and unsupervised machine learning.

PyPy is not a package; it is an alternative implementation of Python 2.7.8 that supports most of the commonly used Python standard packages (unfortunately, NumPy is currently not fully supported). As an advantage, it offers enhanced speed and memory handling. Thus, it is very useful for heavy duty operations on large chunks of data and it should be part of your big data handling strategies.

The Scikit-learn toy dataset is embedded in the Scikit-learn package. Such datasets can easily be directly loaded into Python by the import command, and they don't require any download from any external Internet repository. Some examples of this type of dataset are the Iris, Boston, and Digits datasets, to name the principal ones mentioned in uncountable publications and books, and a few other classic ones for classification and regression.

Structured in a dictionary-like object, besides the features and target variables, they offer complete descriptions and contextualization of the data itself.

For instance, to load the Iris dataset, enter the following commands:

In: from sklearn import datasets
In: iris = datasets.load_iris()

After loading, we can explore the data description and understand how the features and targets are stored. Basically, all Scikit-learn datasets present the following methods:

Now, let's just try to implement them (no output is reported, but the print commands will provide you with plenty of information):

In: print iris.DESCR
In: print iris.data
In: print iris.data.shape
In: print iris.feature_names
In: print iris.target
In: print iris.target.shape
In: print iris.target_names

Now, you should know something more about the dataset—about how many examples and variables are present and what their names are.

Notice that the main data structures that are enclosed in the iris object are the two arrays, data and target:

In: print type(iris.data)
Out: <type 'numpy.ndarray'>

Iris.data offers the numeric values of the variables named sepal length, sepal width, petal length, and petal width arranged in a matrix form (150,4), where 150 is the number of observations and 4 is the number of features. The order of the variables is the order presented in iris.feature_names.

Iris.target is a vector of integer values, where each number represents a distinct class (refer to the content of target_names; each class name is related to its index number and setosa, which is the zero element of the list, is represented as 0 in the target vector).

The Iris flower dataset was first used in 1936 by Ronald Fisher, who was one of the fathers of modern statistical analysis, in order to demonstrate the functionality of linear discriminant analysis on a small set of empirically verifiable examples (each of the 150 data points represented iris flowers). These examples were arranged into tree balanced species classes (each class consisted of one-third of the examples) and were provided with four metric descriptive variables that, when combined, were able to separate the classes.

The advantage of using such a dataset is that it is very easy to load, handle, and explore for different purposes, from supervised learning to graphical representation. Modeling activities take almost no time on any computer, no matter what its specifications are. Moreover, the relationship between the classes and the role of the explicative variables are well known. So, the task is challenging, but it is not arduous.

For example, let's just observe how classes can be easily separated when you wish to combine at least two of the four available variables by using a scatterplot matrix.

Scatterplot matrices are arranged in a matrix format, whose columns and rows are the dataset variables. The elements of the matrix contain single scatterplots whose x values are determined by the row variable of the matrix and y values by the column variable. The diagonal elements of the matrix may contain a distribution histogram or some other univariate representation of the variable at the same time in its row and column.

The pandas library offers an off-the-shelf function to quickly make up scatterplot matrices and start exploring relationship and distributions between the quantitative variables in a dataset.

In:
import pandas as pd
import numpy as np
In: colors = list()
In: palette = {0: "red", 1: "green", 2: "blue"}
In:
for c in np.nditer(iris.target): colors.append(palette[int(c)])
    # using the palette dictionary, we convert
    # each numeric class into a color string
In: dataframe = pd.DataFrame(iris.data, columns=iris.feature_names)
In: scatterplot = pd.scatter_matrix(dataframe, alpha=0.3, figsize=(10, 10), diagonal='hist', color=colors, marker='o', grid=True)

We encourage you to expriment a lot with this dataset and with similar ones before you work on other complex real data, because the advantage of focusing on an accessible, non-trivial data problem is that it can help you to quickly build your foundations on data science.

After a while, anyway, though useful and interesting for your learning activities, toy datasets will start limiting the variety of different experimentations that you can achieve. In spite of the insight provided, in order to progress, you'll need to gain access to complex and realistic data science topics. We will, therefore, have to resort to some external data.

The second type of example dataset that we will present can be downloaded directly from the machine learning dataset repository, or from the LIBSVM data website. Contrary to the previous dataset, in this case, you will need to have access to the Internet.

First of all, mldata.org is a public repository for machine learning datasets that is hosted by the TU Berlin University and supported by Pattern Analysis, Statistical Modelling, and Computational Learning (PASCAL), a network funded by the European Union.

For example, if you need to download all the data related to earthquakes since 1972 as reported by the United States Geological Survey, in order to analyze the data to search for predictive patterns you will find the data repository at http://mldata.org/repository/data/viewslug/global-earthquakes/ (here, you will find a detailed description of the data).

Note that the directory that contains the dataset is global-earthquakes; you can directly obtain the data using the following commands:

In: from sklearn.datasets import fetch_mldata
In: earthquakes = fetch_mldata('global-earthquakes')
In: print earthquakes.data
In: print earthquakes.data.shape
Out: (59209L, 4L)

As in the case of the Scikit-learn package toy dataset, the obtained object is a complex dictionary-like structure, where your predictive variables are earthquakes.data and your target to be predicted is earthquakes.target. This being the real data, in this case, you will have quite a lot of examples and just a few variables available.

LIBSVM Data (http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/) is a page-gathering data from many other collections. It offers different regression, binary, and multilabel classification datasets stored in the LIBSVM format. This repository is quite interesting if you wish to experiment with the support vector machine's algorithm.

If you want to load a dataset, first go to the page where you wish to visualize the data. In this case, visit http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary/a1a and take down the address. Then, you can proceed by performing a direct download:

In: import urllib2
In: target_page = 'http://www.csie.ntu.edu.tw/~cjlin/libsvmtools/datasets/binary/a1a'
In: a2a = urllib2.urlopen(target_page)
In: from sklearn.datasets import load_svmlight_file
In: X_train, y_train = load_svmlight_file(a2a)
In: print X_train.shape, y_train.shape
Out: (2265, 119) (2265L,)

In return, you will get two single objects: a set of training examples in a sparse matrix format and an array of responses.

Sometimes, you may have to download the datasets directly from their repository using a web browser or a wget command.

If you have already downloaded and unpacked the data (if necessary) into your working directory, the simplest way to load your data and start working is offered by the NumPy and the pandas library with their respective loadtxt and read_csv functions.

For instance, if you intend to analyze the Boston housing data and use the version present at http://mldata.org/repository/data/viewslug/regression-datasets-housing/, you first have to download the regression-datasets-housing.csv file in your local directory.

Since the variables in the dataset are all numeric (13 continuous and one binary), the fastest way to load and start using it is by trying out the NumPy function loadtxt and directly loading all the data into an array.

Even in real-life datasets, you will often find mixed types of variables, which can be addressed by pandas.read_table or pandas.read_csv. Data can then be extracted by the values method; loadtxt can save a lot of memory if your data is already numeric since it does not require any in-memory duplication.

In: housing = np.loadtxt('regression-datasets-housing.csv',delimiter=',')
In: print type(housing)
Out: <type 'numpy.ndarray'>
In: print housing.shape
Out:(506L, 14L)

The loadtxt function expects, by default, tabulation as a separator between the values on a file. If the separator is a colon (,) or a semi-colon(;), you have to explicit it using the parameter delimiter.

>>>  import numpy as np
>>> type(np.loadtxt)
<type 'function'>
>>> help(np.loadtxt)

Help on function loadtxt in module numpy.lib.npyio.

Another important default parameter is dtype, which is set to float.

If you need to determinate a different type (for example, an int), you have to declare it beforehand.

For instance, if you want to convert numeric data to int, use the following code:

In: housing_int = np.loadtxt('regression-datasets-housing.csv',delimiter=',', dtype=int)

Printing the first three elements of the row of the housing and housing_int arrays can help you understand the difference:

In: print housing[0,:3], '\n', housing_int[0,:3]
Out:
[  6.32000000e-03   1.80000000e+01   2.31000000e+00]
[ 0 18  2]

Frequently, though not always the case in our example, the data on files feature in the first line a textual header that contains the name of the variables. In this situation, the parameter that is skipped will point out the row in the loadtxt file from where it will start reading the data. Being the header on row 0 (in Python, counting always starts from 0), parameter skip=1 will save the day and allow you to avoid an error and fail to load your data.

The situation would be slightly different if you were to download the Iris dataset, which is present at http://mldata.org/repository/data/viewslug/datasets-uci-iris/. In fact, this dataset presents a qualitative target variable, class, which is a string that expresses the iris species. Specifically, it's a categorical variable with four levels.

Therefore, if you were to use the loadtxt function, you will get a value error due to the fact that an array must have all its elements of the same type. The variable class is string, whereas the other variables are constituted of floating point values.

How to proceed? The pandas library offers the solution, thanks to its DataFrame data structure that can easily handle datasets in a matrix form (row per columns) that is made up of different types of variables.

First of all, just download the datasets-uci-iris.csv file and have it saved in your local directory.

At this point, using pandas' read_csv is quite straightforward:

In: iris_filename = 'datasets-uci-iris.csv'
In: iris = pd.read_csv(iris_filename, sep=',', decimal='.', header=None, names= ['sepal_length', 'sepal_width', 'petal_length', 'petal_width', 'target'])
In: print type(iris)
Out: <class
 'pandas.core.frame.DataFrame'>

Apart from the filename, you can specify the separator (sep), the way the decimal points are expressed (decimal), whether there is a header (in this case, header=None; normally, if you have a header, then header=0), and the name of the variable—where there is one (you can use a list; otherwise, pandas will provide some automatic naming).

If, at this point, you need to convert the pandas DataFrame into a couple of NumPy arrays that contain the data and target values, this can be easily done in a couple of commands:

In: iris_data = iris.values[:,:4]
In: iris_target, iris_target_labels = pd.factorize(iris.target)
In: print iris_data.shape, iris_target.shape
Out: (150L, 4L) (150L,)

As a last learning resource, Scikit-learn also offers the possibility to quickly create synthetic datasets for regression, binary and multilabel classification, cluster analysis, and dimensionality reduction.

The main advantage of recurring to synthetic data lies in its instantaneous creation in the working memory of your Python console. It is, therefore, possible to create bigger data examples without having to engage in long downloading sessions from the Internet (and saving a lot of stuff on your disk).

For example, you may need to work on a million example classification problem:

In: from sklearn import datasets # We just import the "datasets" module
In: X,y = datasets.make_classification(n_samples=10**6, n_features=10, random_state=101)
In: print X.shape,  y.shape
Out: (1000000L, 10L) (1000000L,)

After importing just the datasets module, we ask, using the make_classification command, for 1 million examples (the n_samples parameter) and 10 useful features (n_features). The random_state should be 101, so we can be assured that we can replicate the same datasets at a different time and in a different machine.

For instance, you can type the following command:

$> datasets.make_classification(1, n_features=4, random_state=101)

This will always give you the following output:

(array([[-3.31994186, -2.39469384, -2.35882002,  1.40145585]]), array([0]))

No matter what the computer and the specific situation is, random_state assures deterministic results that make your experimentations perfectly replicable.

Defining the random_state parameter using a specific integer number (in this case 101, but it may be any number that you prefer or find useful) allows the easy replication of the same dataset on your machine, the way it is set up, on different operating systems, and on different machines.

By the way, did it take too long?

On a i3-2330M CPU @ 2.20GHz machine, it takes:

In: %timeit X,y = datasets.make_classification(n_samples=10**6, n_features=10, random_state=101)
Out: 1 loops, best of 3: 2.17 s per loop

If it doesn't seem so also on your machine and if you are ready, having set up and tested everything up to this point, we can start our data science journey.