Applied Supervised Learning with R: Use machine learning libraries of R to build models that solve business problems and predict future trends

In this exercise, we will write an R script to download the Portuguese Bank Direct Campaign dataset from UCI Machine Learning Repository and extract the content of the ZIP file in a given folder using the unzip function.

Preform these steps to complete the exercise:

CSV files are the most common type of data storage and exchange formats for structured data. R provides a method called read.csv() for reading data from a CSV file. It will read the data into a data.frame (more about it in the next section). There are many arguments that the method takes; the two required arguments are a path to the filename and sep, which specifies the character that separates the column values. The summary() method describes the six summary statistics, min, first quartile, median, mean, third quartile, and max.

In the following exercise, we'll read a CSV file and summarize its column.

In this exercise, we will read the previously extracted CSV file and use the summary function to print the min, max, mean, median, 1st quartile, and 3rd quartile values of numeric variables and count the categories of the categorical variable.

Carry out these steps to read a CSV file and later summarize its columns:

JSON is the next most commonly used data format for sharing and storing data. It is unlike CSV files, which only deal with rows and columns of data where each has a definite number of columns. For example, in the e-commerce data of the customers, each row could be representing a customer with their information stored in separate columns. For a customer, if a column has no value, the field is stored as NULL.

JSON provides an added flexibility of having a varying number of fields for each customer. This type of flexibility relieves the developer from the burden of maintaining a schema as we have in traditional relational databases, wherein the same customer data might be spread across multiple tables to optimize for storage and querying time.

JSON is more of a key-value store type of storage, where all we care about is the keys (such as the name, age, and DOB) and their corresponding values. While this sounds flexible, proper care has to be taken, otherwise manageability might at times, go out of control. Fortunately, with the advent of big data technologies in recent days, many document stores (a subclass of the key-value store), popularly also known as NoSQL databases, are available for storing, retrieving, and processing data in such formats.

In the following exercise, the JSON file has data for cardamom (spices and condiments) cultivation district-wise in Tamil Nadu, India, for the year 2015-16. The keys include area (hectare), production (in quintals), and productivity (average yield per hectare).

The jsonlite package provides an implementation to read and convert a JSON file into DataFrame, which makes the analysis simpler. The fromJSON method reads a JSON file, and if the flatten argument in the fromJSON function is set to TRUE, it gives a DataFrame.

In this exercise, we will read a JSON file and store the data in the DataFrame.

Perform the following steps to complete the exercise:

Unstructured data is the language of the web. All the social media, blogs, web pages, and many other sources of information are textual and untidy to extract any meaningful information. An increasing amount of research work is coming out from the Natural Language Processing (NLP) field, wherein computers are becoming better in understanding not only the meaning of the word but also the context in which it's used in a sentence. The rise of computer chatbot, which responds to a human query, is the most sophisticated form of understanding textual information.

In R, we will use the tm text mining package to show how to read, process, and retrieve meaningful information from text data. We will use a small sample of the Amazon Food Review dataset in Kaggle (https://www.kaggle.com/snap/amazon-fine-food-reviews) for the exercise in this section.

In the tm package, collections of text documents are called Corpus. One implementation of Corpus in the tm package is VCorpus (volatile corpus). Volatile corpus is named after the fact that it's stored in-memory for fast processing. To check the metadata information of the VCorpus object, we can use the inspect() method. The following exercise uses the lapply method for looping through the first two reviews and casting the text as a character. You will learn more about the apply family of function in the The Apply Family of Functions section.

In this exercise, we will read a CSV file with the text column and store the data in VCorpus.

Perform the following steps to complete the exercise:

We will revisit the review_corpus dataset in a later section to show how to convert the unstructured textual information to structured tabular data.

Apart from CSV, Text, and JSON, there are numerous other data formats depending upon the source of data and its usage. R has a rich collection of libraries that helps with many formats. R can import not only the standard formats (apart from the previous three) such as HTML tables and XML but also formats specific to an analytical tool such as SAS and SPSS. This democratization led to a significant migration of industry experts who were earlier working in the propriety tools (costly and often found with only the large corporations) to open source analytical programming languages such as R and Python.

In this activity, we will create a R Markdown file to read a CSV file and print a small summary of the data in a word file:

Perform the following steps to complete the activity:

Vector is the most fundamental of all the data structures, and the values are stored in a 1-D array. Vector is the most suitable for a single variable with a series of values. In Exercise 3, Reading a JSON File and Storing the Data in DataFrame, refer to step 4 where we assigned a DataFrame its column names and concatenated using the c() method, as shown here:

c_names <- c("S.No","District","Area","Production","PTY")

We can extract the second value in the vector by specifying the index in square brackets next to the vector name. Let's review the following code where we subset the value in the second index:

c_names[2]

The output is as follows:

## [1] "District"

The collection of string concatenated with the c() method is a vector. It can store a homogenous collection of characters, integers, or floating point values. While trying to store an integer with character, an implicit type cast will happen, which will convert all the values to character.

Matrix is the higher dimension data structure used for storing n-dimensional data. It is suitable for storing tabular data. Similar to vector, the matrix also allows only homogenous collection of data in its rows and columns.

The following code generates 16 random numbers drawn from a binomial distribution with a parameter, number of trials (size) = 100, and success probability equal to 0.4. The rbinom() method in R is useful for generating such random numbers:

r_numbers <- rbinom(n = 16, size = 100, prob = 0.4)

Now, to store r_number as a matrix, use the following command:

matrix(r_numbers, nrow = 4, ncol = 4)

The output is as follows:

##      [,1] [,2] [,3] [,4]
## [1,]   48   39   37   39
## [2,]   34   41   32   38
## [3,]   40   34   42   46
## [4,]   37   42   36   44

Let's extend the text mining example we took in Exercise 4, Reading a CSV File with Text Column and Storing the Data in VCorpus, to understand the usage of matrix in text mining.

Consider the following two reviews. Use the lapply to type cast the first review to as.character and print:

lapply(review_corpus[1:2], as.character)

The output is as follows:

## $'1'
## [1] "I have bought several of the Vitality canned dog food products and have found them all to be of good quality. The product looks more like a stew than a processed meat, and it smells better. My Labrador is finicky, and she appreciates this product better than  most."

## $'2'
## [1] "Product arrived labeled as Jumbo Salted Peanuts...the peanuts were actually small sized unsalted. Not sure if this was an error or if the vendor intended to represent the product as \"Jumbo\".

Now, in the following exercise, we will transform the data to remove stopwords, whitespaces, and punctuations from these two paragraphs. We will then perform stemming (both looking and looked will be reduced to look). Also, for consistency, convert all the text into lowercase.

In this exercise, we will perform the transformation on the data to make it available for further analysis.

Perform the following steps to complete the exercise:

While vector and matrix both are useful structures to be used in various computations in a program, it might not be sufficient for storing a real-world dataset, which most often contains data of mix types, like a customer table in CRM application has the customer name and age together in two columns. The list offers a structure to allow for storing two different types of data together.

In the following exercise, along with generating 16 random numbers, we have used the sample() method to generate 16 characters from the English alphabet. The list method stores both the integers and characters together.

In this exercise, we will use the list method to store randomly generated numbers and characters. The random numbers will be generated using the rbinom function, and the random characters will be selected from English alphabets A-Z.

Perform the following steps to complete the exercise:

In this activity, you will create two matrices and retrieve a few values using the index of the matrix. You will also perform operations such as multiplication and subtraction.

Perform the following steps to complete the activity:

Let's revisit step 6 of Exercise 6, Using the List Method for Storing Integers and Characters Together, where we discussed the integrity check when we attempted to store two unequal length vectors in a list and will see how DataFrame handles it differently. We will, once again, generate random numbers (r_numbers) and random characters (r_characters).

Perform the following steps to complete the exercise:

Accessing any particular row and column in the DataFrame is similar to the matrix. We will show many tricks and techniques to best use the power of indexing in the DataFrame, which also includes some of the filtering options.

Every row in a DataFrame is a result of the tightly coupled collection of columns. Each column clearly defines the relationship each row of data has with every other one. If there is no corresponding value available in a column, it will be filled with NA. For example, a customer in a CRM application might not have filled their marital status, whereas a few other customers filled it. So, it becomes essential during application design to specify which columns are mandatory and which are optional.

With the growing adaption of DataFrame came a time when its limitations started to surface. Particularly with large datasets, DataFrame performs poorly. In the complex analysis, we often create many intermediate DataFrames to store the results. However, R is built on an in-memory computation architecture, and it heavily depends on RAM. Unlike disk space, RAM is limited to either 4 or 8 GB in many standard desktops and laptops. DataFrame is not built efficiently to manage the memory during the computation, which often results in out of memory error, especially when working with large datasets.

In order to handle this issue, data.table inherited the data.frame functionality and offers fast and memory-efficient version for the following task on top of it:

Efficient memory management makes the development fast and reduces the latency between operations. The following exercise shows the significant difference data.table makes in computation time as compared to data.frame. First, we read the complete Amazon Food Review dataset, which is close to 286 MB and contains half a million records (this is quite a big dataset for R), using the fread() method, which is one of the fast reading methods from data.table.

In this exercise, we will only show file read operations. You are encouraged to test the other functionalities (https://cran.r-project.org/web/packages/data.table/vignettes/datatable-intro.html) and compare the data table capabilities over DataFrame.

Perform the following steps to complete the exercise:

Observe that 3.78 seconds elapsed for reading it through the fread() method, while the read.csv function took 4.91 seconds. The execution speed is almost 30% faster. As the size of the data increasing, this difference is even more significant.

In the previous output, the user time is the time spent by the current R session, and system time is the time spent by the operating system to complete the process. It's possible that you may get a different value after executing the system.time method even if you use the same dataset. It depends a lot on how busy your CPU was at the time of running the method. However, we should read the output of the system.time method relative to the comparison we are carrying out and not relative to the absolute values.

When the size of the dataset is too large, we have too many intermediate operations to get to the final output. However, keep in mind that data.table is not the magic wand that allows us to deal with a dataset of any size in R. The size of RAM still plays a significant role, and data.table is no substitute for distributed and parallel processing big data systems. However, even for the smaller dataset, the usage of data.table has shown much better performance than data.frames.

As the name suggests, it combines two or more vector, matrix, DataFrame, or table by column. cbind is useful when we have more than one vector, matrix, or DataFrame that need to be combined into one for analysis or visualization. The output of cbind varies based on the input data. The following exercise provides a few examples of cbind, which combines two vectors.

In this exercise, we will implement the cbind function to combine two DataFrame objects.

Perform the following steps to complete the exercise:

The error means that the lengths of r_numbers and r_characters are not same (16 and 18, respectively). Note that the cbind() method, unlike as.data.frame(), doesn't throw an error. Instead, it automatically performs what is known as Recycling, and the vector of shorter length gets recycled. In the output, the r_numbers 38 and 48 are recycled from the top to fill the 17th and 18th index.

Consider that we write the following command instead:

cbind(as.data.frame(r_numbers), as.data.frame(r_characters))

It will now throw an error as we had shown earlier in the DataFrame section:

Error in data.frame(..., check.names = FALSE) : 
  arguments imply differing number of rows: 16, 18

One needs to be careful by always checking for the dimensions and the class of data. Otherwise, it may lead to unwanted results. When we give two vectors, it creates a matrix by default on doing a cbind.

rbind is like cbind, but it combines by row instead of column. For rbind to work, the number of columns should be equal in both the DataFrames. It is useful in cases when we want to append an additional set of observations with an existing dataset where all the columns of the original dataset are the same and are in the same order. Let's explore rbind in the following exercise.

In this exercise, we will combine two DataFrames using the rbind function.

Perform the following steps to complete the exercise:

From the last step, observe that the rbind function concatenates (binds) the r_numbers and r_characters as two rows of data, unlike cbind, where it was stacked in two columns. Except for the output, all the other rules of cbind apply to rbind as well.

The merge() function in R is particularly useful when there is more than one DataFrame to join using a common column (what we call a primary key in the database world). Merge has two different implementations for the DataFrame and data table, which behave mostly in the same way.

In this exercise, we will generate two DataFrames, that is, df_one and df_two, such that the r_numbers column uniquely identifies each row in each of the DataFrame.

Perform the following steps to complete the exercise:

First DataFrame

Second DataFrame

Once we create the df_one and df_two DataFrames using the cbind() function, we are ready to perform some merge (will use the word JOIN, which means the same as merge()).

Now, let's see how different type of joins give different results.

In the world of databases, JOINs are used to combine two or more than two tables using a common primary key. In databases, we use Structured Query Language (SQL) to perform the JOINs. In R, the merge() function helps us with the same functionality as SQL offers in databases. Also, instead of tables, we have DataFrames here, which is again a table with rows and columns of data.

In Exercise 11, Exploring the merge Function, we created two DataFrames: df_one and df_ two. We will now join the two DataFrames using Inner Join. Observe that only the value 26 (row number 7) in the r_numbers column is common between the two DataFrames, where the corresponding character in the r_characters column is R in df_one and character O in df_two. In the output, X corresponds to the df_one DataFrame and Y correspond to the df_two DataFrame.

To merge the df_one and df_two DataFrames using the r_numbers column, use the following command:

merge(df_one, df_two, by = "r_numbers")
##   r_numbers r_characters.x r_characters.y
## 1        26              R              O

Left Join gives all the values of df_one in the r_numbers column and adds <NA> as a value wherever the corresponding value in df_two is not found. For example, for r_number = 2, there is no value in df_two, whereas for r_number = 26, values in df_one and df_two, for the r_characters column is R and O, respectively.

To merge the df_one and df_two DataFrames using the r_numbers column, use the following command:

merge(df_one, df_two, by = "r_numbers", all.x = TRUE)
##    r_numbers r_characters.x r_characters.y
## 1          2              K           <NA>
## 2          4              R           <NA>
## 3          8              W           <NA>
## 4          9              J           <NA>
## 5         10              Q           <NA>
## 6         13              T           <NA>
## 7         16              H           <NA>
## 8         20              F           <NA>
## 9         25              J           <NA>
## 10        26              R              O

Right Joins works just like Left Join, except for that the values in the r_character columns of df_one are <NA> wherever a match is not found. Again, r_numbers = 26 is the only match.

To merge the df_one and df_two DataFrames using the r_numbers column, use the following command:

merge(df_one, df_two, by = "r_numbers", all.y = TRUE)
##    r_numbers r_characters.x r_characters.y
## 1          3           <NA>              D
## 2          5           <NA>              Q
## 3         12           <NA>              X
## 4         17           <NA>              L
## 5         18           <NA>              J
## 6         19           <NA>              Q
## 7         21           <NA>              H
## 8         26              R              O
## 9         29           <NA>              D
## 10        30           <NA>              Q

Unlike Left and Right Join, Full Join gives all the unique values of the r_numbers column from both the DataFrames and adds <NA> in the r_characters column from the respective DataFrame. Observe that only the r_number = 26 row has values from both the DataFrame.

To merge the df_one and df_two DataFrames using the r_numbers column, use the following command:

merge(df_one, df_two, by = "r_numbers", all = TRUE)
##    r_numbers r_characters.x r_characters.y
## 1          2              K           <NA>
## 2          3           <NA>              D
## 3          4              R           <NA>
## 4          5           <NA>              Q
## 5          8              W           <NA>
## 6          9              J           <NA>
## 7         10              Q           <NA>
## 8         12           <NA>              X
## 9         13              T           <NA>
## 10        16              H           <NA>
## 11        17           <NA>              L
## 12        18           <NA>              J
## 13        19           <NA>              Q
…

Data is known to be in a wide format if each subject has only a single row, with each measurement present as a different variable or column. Similarly, it is a long format if each measurement has a single observation (thus, multiple rows per subject). The reshape function is used often to convert between wide and long formats for a variety of operations to make the data useful for computation or analysis. In many visualizations, we use reshape() to convert wide format to long and vice versa.

We will use the Iris dataset. This dataset contains variables named Sepal.Length, Sepal.Width, Petal.Length, and Petal.Width, whose measurements are given in centimeters, for 50 flowers from each of 3 species of Iris, namely setosa, versicolor, and virginica.

In this exercise, we will explore the reshape function.

Perform the following steps to complete the exercise:

This warning means there were multiple values for Type 1 and Type 2 for the three species, so the reshape has picked the first occurrence of each of the species. In this case, the 1, 51, and 101 row numbers. We will now see how we could handle this transformation better in the aggregate function.

Aggregation is a useful method for computing statistics such as count, averages, standard deviations, and quartiles, and it also allows for writing a custom function. In the following code, the formula (formula is a name of the data structure in R, not a mathematical equation) for each Iris species computes the mean of the numeric measures sepal and petal width and length. The first of the aggregate function argument is a formula that takes species and all the other measurements to compute the mean from all the observations.

aggregate(formula =. ~ Species, data = iris, FUN = mean)

The output of the previous command is as follows:

##      Species Sepal.Length Sepal.Width Petal.Length Petal.Width
## 1     setosa        5.006       3.428        1.462       0.246
## 2 versicolor        5.936       2.770        4.260       1.326
## 3  virginica        6.588       2.974        5.552       2.026

The apply() function takes an array, including a matrix, as input and returns a vector, array, or list of values obtained by applying a function to margins of an array or matrix.

In this exercise, we will count the number of vowels in each column of a 100 x 100 matrix of random letters from the English alphabet. The MARGIN = 1 function will scan each row, and MARGIN = 2 will specify the column. The same function will the count vowels in each row.

Perform the following steps to complete the exercise:

The lapply function looks similar to apply(), with a difference that it takes input as a list and returns a list as output. After rewriting our previous example in the following exercise, the output of class function shows that the output is a list.

In this exercise, we will take a list of vectors and count the number of vowels.

Perform the following steps to complete the exercise:

The sapply function is just a wrapper on the lapply function, where the output is a vector or matrix instead of a list. In the following code, observe the type of the output after applying sapply difference. The output returns a vector of integers, as we can check with the class() function:

sapply(r_characters, c_vowel)
##  a  b 
## 19 10

To print the class of the output, use the following command:

out_vector <- sapply(r_characters, c_vowel)
class(out_vector)

The output of the previous command is as follows:

## [1] "integer"

Apply a function to each cell of a ragged array, that is, to each (non-empty) group of values given by a unique combination of the levels of certain factors. The tapply function is quite useful when it comes to working on a subset level of data. For example, in our aggregate function, if we were to get an aggregate like standard deviation for the type of Iris species, we could use tapply. The following code shows how to use the tapply function:

First, calculate the standard deviation of sepal length for each Iris species:

tapply(iris$Sepal.Length, iris$Species,sd)

The output is as follows:

##     setosa versicolor  virginica 
##  0.3524897  0.5161711  0.6358796

Next, calculate the standard deviation of sepal width for each of the Iris species:

tapply(iris$Sepal.Width, iris$Species,sd)

The output of the previous command is as follows:

##     setosa versicolor  virginica 
##  0.3790644  0.3137983  0.3224966

Now, let's explore some popular and useful R packages that might be of value while building complex data processing methods, machine learning models, or data visualization.

The dplyr package helps in the most common data manipulation challenges through five different methods, namely, mutate(), select(), filter(), summarise(), and arrange(). Let's revisit our direct marketing campaigns (phone calls) of a Portuguese banking institution dataset from UCI Machine Learning Repository to test out all these methods.

The %>% symbol in the following exercise is called chain operator. The output of the one operation is sent to the next one without explicitly creating a new variable. Such a chaining operation is storage efficient and makes the readability of the code easy.

In this exercise, we are interested in knowing the average bank balance of people doing blue-collar jobs by their marital status. Use the functions from the dplyr package to get the answer.

Perform the following steps to complete the exercise:

Much of complex analysis is done with ease. Note that the mutate() method helps in creating custom columns with certain calculation or logic.

The tidyr package has three essential functions—gather(), separate(), and spread()—for cleaning messy data.

The gather() function converts wide DataFrame to long by taking multiple columns and gathering them into key-value pairs.

In this exercise, we will explore the tidyr package and the functions associated with it.

Perform the following steps to complete the exercise:

This activity will make you accustomed to selecting all numeric fields from the bank data and produce the summary statistics on numeric variables.

Perform the following steps to complete the activity:

What we saw with the apply functions could be done through the plyr package on a much bigger scale and robustness. The plyr package provides the ability to split the dataset into subsets, apply a common function to each subset, and combine the results into a single output. The advantage of using plyr over the apply function is features like the following:

All the function names in plyr are clearly defined based on input and output. For example, if an input is a DataFrame and output is list, the function name would be dlply.

The following figure from the The Split-Apply-Combine Strategy for Data Analysis paper displays all the different plyr functions:

Figure 1.7: Functions in the plyr packages

The _ means the output will be discarded.

In this exercise, we will see how split-apply-combine makes things simple with the flexibility of controlling the input and output.

Perform the following steps to complete the exercise:

In steps 5, 6, and 7, observe how we created a list, array, and data as an output for DataFrame input. All we must do is use a different function from plyr. This makes it easy to type cast between many possible combinations.

The caret package is particularly useful for building a predictive model, and it provides a structure for seamlessly following the entire process of building a predictive model. Starting from splitting data to training and testing dataset and variable importance estimation, we will extensively use the caret package in our chapters on regression and classification. In summary, caret provides tools for:

We will revisit the caret package with examples in Chapter 4, Regression, and Chapter 5, Classification.

A scatterplot is a type of plot or mathematical diagram using Cartesian coordinates to display values for typically two variables for a set of data. If the points are color-coded, an additional variable can be displayed.

It is the most common type of chart and is extremely useful in spotting patterns in the data, especially between two variables. We will use our bank data again to do some EDA. Let's use the Portuguese bank direct campaign dataset for the visualizations:

df_bank_detail <- read.csv("bank-full.csv", sep = ';')

ggplot works in a layered way of stacking different elements of the plot. In the following example of this section, in the first layer, we provide the data to the ggplot() method and then map it with aesthetic details like x and y-axis, in the example, the age and balance values, respectively. Finally, to be able to identify some reasoning associated with few high bank balances, we added a color based on the type of job.

Execute the following command to plot the scatterplot of age and balance:

ggplot(data = df_bank_detail) +
  geom_point(mapping = aes(x = age, y = balance, color = job))

Figure 1.8: Scatterplot of age and balance.

From Figure 1.8, the distribution of bank balance with age looks much normal, with middle age showing a high bank balance whereas youngsters and old people are on the lower side of the spectrum.

Interestingly, some outlier values seem to be coming from management and retired professionals.

In data visualization, it's always tempting to see a graph and jump to a conclusion. A data visual is for consuming the data better and not for drawing causal inference. Usually, an interpretation by an analyst is always vetted by a business. Graphs that are aesthetically pleasing often tempt you to put it into presentation deck. So, next time a beautiful chart gets into your presentation deck, carefully analyze what you are going to say.

In this section, we will draw three scatter plots in a single plot between age and balance split by marital status (one for each single, divorced, and married individuals).

Now, you could split the distribution by marital status. The patterns seem to be consistent among the single, married, and divorced individuals. We used a method called facet_wrap() as the third layer in ggplot. It takes a marital variable as a formula:

ggplot(data = df_bank_detail) +
  geom_point(mapping = aes(x = age, y = balance, color = job)) +
  facet_wrap(~ marital, nrow = 1)

Figure 1.9: Scatter plot between age and balance split by marital status