Resources for Statistics/Machine Learning That Are Good For Programmers

Statistics are coming up more and in programming situations across all fields of study. Knowing a bit of stats can really help you solve problems with big data sets and enable you to create some interesting features for products. Especially with all the large data sets being created by social media and other online services knowing some statistics can help immensely when trying to create innovative products.

The field of machien learning has been taking off as well, and it being used across all academic disciplines as a way to learn more about complex systems. The past semester I have been doing some work with machine learning and here are some of the resources I’ve found helpful in studies.

Books

Free Courses

Blogs/ Websites

 

There are loads of more resources out there, but these are the ones that have helped me out the most the past few months.

 

Finding new Ideas from Data Visualization

I came a across a nice article that goes over some things to keep in mind when dealing with scientific data and the importance of visualization when programming with large data sets. It got me thinking about how visualizing data can really change your perspective of a certain problem that you are trying to solve. With large data sets it can daunting trying to find some sort of form within them, while at the same time trying to find some sort of conclusion about the data.

Some of the things that Vince Buffalo stressed seemed kind of obvious at first, but the ideas he was presenting got me thinking about how important it is for humans to have some sort of visualization to better understand a concept. By visualizing the data you get a much better perspective on the shape of the date, and it is much easier to interpret than pure numbers or tables. YOu can also view multi variables of data in one visualization and see how variables interact with each other. When working with data the key thing is to understand the data, and visualizations can help you learn how the variables in your data are interacting and give insight into the relationship between them.

Right now I am collecting data on the Chicago Public Transit systems L trains using there API. And found some interesting things by visualizing the data. Using R I imported the CTA’s predicted arrival times for each of the stations on one of there lines. I then found the average value for the wait time at each stop. THe results I got surprised and gave me some insight into what type of factors affect train arrival times in the city. For instance, at the downtown stops where the stations are very close there was some weird patterns for the wait times. All of those stops are very close together compared to most other stops, but for some reason certain stops would have much higher average wait times, even though the stops were right next to each other.

Later on as I am trying to make predictinos on arrival times and the wait times for the CTA, I will definitely take into the population density and factors that may affect the scheduling due to the trains being downtown.

The Monroe stop stood out in the data because the two stops adjacent had much shorter wait times, and the distance between the stops is maybe 1 or 2 minutes while riding. There was also a large difference in wait times depending on which direction the trains were going. This got me thinking about how population density would affect how efficient trains would run, and also the ways the CTA may plan out the frequency of certain train lines depending on population density. For most of the other stops the wait times reflected the distance between the stops, but this trend seemed to change when the stops were in the downtown area. I will write more on this later, but it just goes to show how visualizing data can help you understand patterns in data, and lead you onto new ideas about why the data is taking a certain type of form.

Linear Models In R

Now that you have some data imported into R and understand some of the basic features for summarizes and viewing that data, you can start to dig into some of the features that make R truly standout as a programming language for statistics. The nature of the babies data makes it an excellent choice for utilizing some of the regression features that R offers.

Basically we will want to be predicting numbers using a model we have created with a training set of data. You have one varaibles that you would like to predict and then there are multiple explanatory variables you are using to predict that find value. This post will being going over simple univariate regression, but R is perfectly capable of multivariate linear regression as well as

Basics with Regression

  • The main goals of regression analysis
  • Predicting
  • Predicting housing prices using characteristics of houses
  • Assessing the effect of a variable on other variables
  • Asses impact of on an industry due to raw material price increases
  •  Give a description to a set of data

When using a linear regression the number one assumption is that the data will some how follow a linear relationship between the input variables/predictors (x) and the response (y). If there are more than one predictor then it is a multivariate regression, otherwise is is merely a simple regression. The response variable must be continuous, but all of the other predictors can be discrete, continuous, or categorical. You may have had to plot a fitted line on a scatter plot at some point in your elementary education, as a way to predict the value of something given that the line goes through the majority of the points plotted.Simple drawing a line by eye is good enough for most day to day things, but if you want accuracy in your predictions and the ability to do it with more complex models some math is going to be used. The general form of a regression line is

y = Θ0 + Θ1(x1)

Θ0 represents the y-intercept of the line and Θ1 the slope.

For simple data it isn’t that hard to just draw a line where it appears to be the best, but for accurate predictions you’ll want to turn towards some math. Now if there are more variables you wanted to keep track of, the equation would get much larger and complex.

To fit a line to a collection of points, most linear models will try and minimize the sum and the least squares of the models residuals. Residuals are the points that are either above or below the line and they can be seen as a way to measure the accuracy of a linear regression being made. If there sums are too big that the line is no being fit very well. Or if the sum is enormous, it may mean that the data isn’t right for linear regression.

For now I will go over how to make a simple one variable regression in R, using the babies.txt data and then afterwards an other model using all of the provided data.

If you have moved away from the initial R prompt here is a script with all of the code we went through in the past post.

You can load it into R by running the while R is running in the same directory as the source file. All of the variables created in the script will be available for your use.


>source("Babies.R")

All of the variables should be loaded into the interpreter.

Creating Linear Models

R makes it easy to create linear models for data sets.

First lets see what the data we have looks like


>head(baby.data)
BirthWeight Gestation Parity Age Height Weight Smoke
1 120 284 0 27 62 100 0
2 113 282 0 33 64 135 0
3 128 279 0 28 64 115 1
5 108 282 0 23 67 125 1
6 136 286 0 25 62 93 0
7 138 244 0 33 62 178 0

First we will have it split up the baby.data data frame from the previous post into a training and test set of data.

# input data frame and seed for random value to split up data frame
# this is the basic structure for creating a function in R
#function to split data frame into two parts for training and testing
> splitdf if (!is.null(seed)) set.seed(seed)
index trainindex trainset testset list(trainset=trainset,testset=testset)
}
> two.sets > summary(two.sets[1])
> summary(two.sets[2])
> train.data > test.data 

Making a linear model for birth weight and gestation period would be interesting. We’ll need to the lm() fucntion that R provides to do that.


# create a linear model to predict birthweight given the gestation period
# arguments are first variable to predict values for then a ~ to denote the variables being used to predict,
# then set data= to the baby.data from earlier
> baby.model > coefficients(baby.model)
# summary info about the new model
> summary(baby.model)

Residuals:

Min 1Q Median 3Q Max
-49.348 -11.065 0.218 10.101 57.704

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) -10.75414 8.53693 -1.26 0.208
Gestation 0.46656 0.03054 15.28 —
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 16.74 on 1172 degrees of freedom
Multiple R-squared: 0.1661, Adjusted R-squared: 0.1654
F-statistic: 233.4 on 1 and 1172 DF, p-value: < 2.2e-16 # to display the coefficients used in the model create > coefficients(baby.model)
(Intercept) Gestation
-10.7541389 0.4665569

Next we will want to create a fit line on the linear model we’ve created. This is done using R’s predict() function. It will generate the predicted values for the linear model we are passing it.


# create a linear fit line for the model using the test data
> lm.fit 

A helpful set of bnumbers when dealing with linear models are the residuals.


# Difference between original labels and lm.fit = residuals
> summary((train.data-lm.fit) - baby.model$residuals)

To see how fit the data is to the linear model you can find the root mean squared error. You do this by taking the sum of all the residual values( data points above or bellow the regression line), then taking it’s square root.


# find the root mean squared error to see the sample standard deviation
> rmsError > rmsError print("rmsError = ")
[1] "rmsError = "
> rmsError
[1] 17.15356
# the RMS error is around 17 so this means on average the prediction will be off by around 17 ounces of the birth weight. So it will be within a pound or so.

As you can see R makes it very easy to create basic linear models for data and then use that data to see how predictions can be made using it. In the next part of this series I will go over incoroperating more variables into the linear model, so predictions can be even better.

Here the the entire script created in this post. Make sure the Babies.R source file is in the same directory as this one if you want to see it run correctly. Or otherwise change the path of the argument for source().


source("Babies.R")
head(baby.data)
# input data frame and seed for random value to split up data frame
splitdf <- function(dataframe, seed=NULL) {
    if (!is.null(seed)) set.seed(seed)
  #take index from 1 to the length of the data frame
    index <- 1:nrow(dataframe)
  #find training set index
    trainindex <- sample(index, trunc(length(index)/2))
  # assign training and test set
    trainset <- dataframe[trainindex, ]
    testset <- dataframe[-trainindex, ]
    list(trainset=trainset,testset=testset)
}

two.sets <- splitdf(baby.data, seed=NULL)

train.data <- two.sets$trainset
test.data <- two.sets$testset

#create a linear model of the baby data trying to predict birthweight giv en other factors
baby.model <- lm(formula = BirthWeight ~ Gestation, data = train.data)

#show coefficients for the model
coefficients(baby.model)

# create a linear fit line for the model using the test data
lm.fit <- predict(baby.model, newdata = test.data)

# Difference between original labels and lm.fit = residuals 
summary((train.data-lm.fit) - baby.model$residuals)

# find the root mean squared error to see the sample standard deviation
rmsError <- sqrt(sum(baby.model$residuals * baby.model$residuals)/length(baby.model$residuals))

print("rmsError = ")
rmsError

# the RMS error is around 16 so this means on average the prediction will be off by around 16 ounces of the birth weight

Helpful Links

Basic Importing and Graphing Data In R

Importing and Analyzing Data

One of the most common ways of importing data into R is using a text file that is delimitated with some sort of character. There are also packages that allow you to input data from a database, or other formats. Once the data is imported as a data frame you can start creating models and graphs.

To go through the examples of this post you can download the data file about baby births.

babies.txt

The file is broken down into columns with a header given at the beginning of the file.

Reading CSV Files

To read in a CSV file as a data frame use the read.csv() function. This will return a data frame with the data from the file. If there is a header for the file then you can import that as well for the column labels. R will try and convert the columns of the file into relevant datatypes. It will work most of the time for numbers, booleans, and strings. To convert dates you will have to the date functions that R provides. You can read about the date functions here


# read in data.csv
# head=TRUE imports header values from the csv file and sets them to column names
# set the seperator using the sep argument
> baby.data <- read.csv(file="data/babies.csv", head=TRUE, sep=" ")

# print out the new data's columns
> colnames(baby.data)
[1] "birth_weight" "gestation"    "parity"       "age"          "height"      
[6] "weight"       "smoke"

# if you want to change the column labels you can pass the colnames() function a vector of labels
> colnames(baby.data) <- c("BirthWeight", "Gestation", "Parity", "Age", "Height", "Weight", "Smoke")

The baby birth data set is pretty large so it’s not the best to print out the entire data frame. There are a few functions that R provides that enable you to get a good idea of what the imported data looks like. Use can use head() and tail() to view the first and last 6 rows of the data frame.


# you can use head() to print out the first 6 entries in the data tail
> head(baby.data)
  birth_weight gestation parity age height weight smoke
1          120       284      0  27     62    100     0
2          113       282      0  33     64    135     0
3          128       279      0  28     64    115     1
4          123        NA      0  36     69    190     0
5          108       282      0  23     67    125     1
6          136       286      0  25     62     93     0
# tail returns the last 6
> tail(baby.data)

# new column names
> head(baby.data)
  Birth Weight Gestation Parity Age Height Weight Smoke
1          120       284      0  27     62    100     0
2          113       282      0  33     64    135     0
3          128       279      0  28     64    115     1
4          123        NA      0  36     69    190     0
5          108       282      0  23     67    125     1
6          136       286      0  25     62     93     0

# you can also view one single column of the data frame by using the $ symbol. It will return a vector containing that column's data
# this is good use to when formatting data
> head(baby.data$Weight)

The summary() function gives you some details about the information in a data frame. The minimum, 1st quartile, median, mean 3rd quartile, and maximum. It also gives you information about how many null values (NA’s) there are. You can then use the na.omit() function to omit the null values from the data frame, so there isn’t any bad data left that can cause bugs later on.


# print out a summary of the imported data
> summary(baby.data)
  Birth Weight     Gestation         Parity            Age       
 Min.   : 55.0   Min.   :148.0   Min.   :0.0000   Min.   :15.00  
 1st Qu.:108.8   1st Qu.:272.0   1st Qu.:0.0000   1st Qu.:23.00  
 Median :120.0   Median :280.0   Median :0.0000   Median :26.00  
 Mean   :119.6   Mean   :279.3   Mean   :0.2549   Mean   :27.26  
 3rd Qu.:131.0   3rd Qu.:288.0   3rd Qu.:1.0000   3rd Qu.:31.00  
 Max.   :176.0   Max.   :353.0   Max.   :1.0000   Max.   :45.00  
                 NA's   :13                       NA's   :2      
     Height          Weight          Smoke       
 Min.   :53.00   Min.   : 87.0   Min.   :0.0000  
 1st Qu.:62.00   1st Qu.:114.8   1st Qu.:0.0000  
 Median :64.00   Median :125.0   Median :0.0000  
 Mean   :64.05   Mean   :128.6   Mean   :0.3948  
 3rd Qu.:66.00   3rd Qu.:139.0   3rd Qu.:1.0000  
 Max.   :72.00   Max.   :250.0   Max.   :1.0000  
 NA's   :22      NA's   :36      NA's   :10

# str will print out the data type for each column in the data frame
# It looks like there are some null so we'll use na.omit() to remove them
> str(baby.data)
'data.frame':  1174 obs. of  7 variables:
 $ BirthWeight: int  120 113 128 108 136 138 132 120 143 140 ...
 $ Gestation  : int  284 282 279 282 286 244 245 289 299 351 ...
 $ Parity     : int  0 0 0 0 0 0 0 0 0 0 ...
 $ Age        : int  27 33 28 23 25 33 23 25 30 27 ...
 $ Height     : int  62 64 64 67 62 62 65 62 66 68 ...
 $ Weight     : int  100 135 115 125 93 178 140 125 136 120 ...
 $ Smoke      : int  0 0 1 1 0 0 0 0 1 0 ...
 - attr(*, "na.action")=Class 'omit'  Named int [1:62] 4 40 43 86 90 94 99 103 111 114 ...
  .. ..- attr(*, "names")= chr [1:62] "4" "40" "43" "86" ...
> baby.data <- na.omit(baby.data) 

Because this data set is reletively large taking out a few rows because of null values is fine and shouldn’t affect results we might try to get from the data.

Graphing

Now that you have some data points loaded into R you can start to to create graphs of the data and see if there are any visible trends that you can find. In the last tutorial you should have installed the ggplot2 package for plotting data. ggplot2 extends R’s graphing functions and allows for graphs to be made layer by layer. It also has better control over the visual aspects of the graphs. ggplot2 creates graphs by layers represented by R functions.

The example below starts off with a ggplot() object with the aesthetics x and y being set to columns in the baby.data data frame . Next, geom_point() creates a scatter plot of the data with the color being represented by the Smoke column in the baby data frame. That means the color of each point will be colored depending on on the Smoke column. The theme is set to black and white with theme_bw(), then scale_color_manual() defines the darkblue color and removes the legend from the data. xlab() and ylab() set the labels for the plot and opts() sets the title. ggsave() saves the file as a pdf to a directory of your choice.


# start off with a ggplot() objhect then add on the other layers to build the graph
> baby.plot<-ggplot(baby.data, aes(x=BirthWeight,y=Gestation,))+
> geom_point(aes(colour=Smoke))+ 
> theme_bw()+
> ylab("Gestation Period In Days")+
> xlab("Birth Weight in Ounces")+
> opts(title="Baby Weight Compared to Gestation Period")+
# fit a line using a linear model (lm)
> stat_smooth(method="lm")
> ggsave(plot=baby.plot, filename="./graphs/babies.pdf",width=8,height=9)

To create separate graphs for the baby data based on the parity of their birth you can use the
facet_wrap() function to break down the data into multiple graphs based on the value of a third column in the data frame.

> baby.plot<-ggplot(baby.data, aes(x=BirthWeight,y=Gestation,))+
> geom_point(aes(colour=Smoke))+ 
> theme_bw()+
> facet_wrap(~Parity, nrow = 6, ncol = 1) + 
> ylab("Gestation Period In Days")+
> xlab("Birth Weight in Ounces")+
> opts(title="Baby Weight Compared to Gestation Period")+
# fit a line using a linear model (lm)
> stat_smooth(method="lm")
> ggsave(plot=baby.plot, filename="./graphs/babies-by-parity.pdf",width=8,height=9)

Graphs made with ggplot2 can be customized in almost any way to fit the type of graph. The docs can be found at the bottom of the page. In the next section of the tutorial I will go over graphing a bit more and take on some simple regressions and clustering.

Part 3: Linear Models In R

Helpful Links

Importing Data in R
PostgreSQL interface for R scripts
ggplot2 Documentation
R Cook Book: Graphing

Introduction the R Programming Language

Introduction to R

R is a programming language and environment for doing computations that involve statistics. The language offers a wide variety of tools that making statistical computations easy to do and takes out much of the boiler-plate code needed to implement various statistical models. It has better support for object orientated programming techniques. There are also strong tools for visualizing data built into R as well as a vast amount of community made packages that extend R’s functionality. I’m going to begin by going over some of the basic features of R and the syntax for some common statistical tasks.

To begin you will need to get an installation of R running on your computer. Most installations come bundled with a GUI editor that highlights R syntax, but most things I will go over can be done using the interactive prompt.

Packages

One great feature of R is the ability to quickly install packages created by the community from the command line. R’s Comprehensive R Archive Network (CRAN) hosts a wide array of packages that extend the functionality of R. You can find a list of available packages here. Installing a package is a simple one line command. Later on in this series, we will be plotting graphs of data using an extension of R’s graphing features called ggplot2. It simplifies some of the steps required for graphing data and allows you to make complex layered graphs. To install the package run


# select a mirror when prompted
> install.packages(“ggplot2”)

R has a good deal of documentation available on the internet as well as a few free books on learning the basics. You can find them here An other helpful feature that R provides the ability to look up documentation on the command line. To look at the documentation on the install.packages() function prepend a ? to the function name and the documentation will be shown. You can also search through the help files for certain terms by using help.search().


> ?install.packages
> help.search("install.packages")

Basics in R

Data Types

R provides a nice selection of data types made for use in statistical computations.

Numbers

In R you can assign numeric values that consist of any real number to a variable. An arrow pointing to the left (<-) is used to assign a value to variable. There is no need to declare the datatype of a variable before assigning it a value.


# assign numbers to two variables
> t <- 45
> u <- 45.543543
# from the interpreter you can print out a variable by typing in the variable name
# print out t and u
> t
[1] 45
> u
[1] 45.54354

Strings

Define a string using single or double quotes.


> y <- "hello world"
> y
[1] "hello world"

Factors

Factors contain values of categorical data that can’t be interpreted as a number. Storing data as a factor makes sure R interprets certain data correctly when creating statistical models. They are an efficient way to store strings, because their values can be stored once in memory and then referenced by a numeric value. More about factors will come up when you start analyzing data.

Vectors

There are three types of vectors in R
-numerical
-character
-logical
They are defined using c().You can refer to an element of a vector using it’s index value, a vector of index values, or a range of index values.


> x <- c(54,43,43,76,21,32)
> y <- c("one", "two", "three", "four")
> z <- c(TRUE, FALSE, TRUE, TRUE, FALSE, FALSE, TRUE)
> x(c(2,5))
# 2nd and 5th index
[1] 43 21
> z[2:5]
#index 2 through 5
[1] FALSE  TRUE  TRUE FALSE

Matrices

Matrices are a large part of computing statistical data in R especially when dealing with regressions and other computations on tabular data. Create a matrix using the matrix() function. Each column in the matrix must be of the same mode type of mode(numeric, character, logical).


#create an 8 x8 matrix of numbers of numbers from 1 to 64
> x <-matrix(1:64, nrow=8, ncol=4)
# create a 4x4 matrix with the columns and rows labels
# byrow=TRUE makes sure the matrix is filled by columns instead of rows
# dimnames is set to a list of of two vectors that will be set as the row and column labels
> y <-matrix(1:16, nrow=2, ncol=2,byrow=TRUE, dimnames=list(c("a", "b"), c("1", "2")))

Lists

Create a list using the list() function. In R lists are like associative arrays with key value pairs. The default key values are integer index values, but they can also be set to strings. Lists can contain any data type and can have mixed types as well.


> a <- c("one", "two", "three", "four")
> b <- "Hello World"
> c <- matrix(1:64, nrow=8, ncol=4)
# create list with variables of mixed data types
> thelist <- list(avector=a, astring=b, amatrix=c)

Data Frames

Data frames are very similar to matrices except the columns of a data frame can have different modes. You can create a data frame by combining multiple vectors of the same length. They can later on be turned into plots and other visualizations. Create one using the data.frame() function. YOu can also uses the names() function to set column labels.


>x <- c(1,2,3,4)
>y <- c("one", "two", "three", "four")
>z <- c(TRUE,FALSE,FALSE,FALSE)
>thedata <- data.frame(x,y,z)
# set the column names of the data frame by passing names() a vector of labels
>names(thedata) <- c("integers","strings","logical") # variable names

  integers strings logical
1        1     one    TRUE
2        2     two   FALSE
3        3   three   FALSE
4        4    four   FALSE

The next section of this tutorial will go over importing and graphing data as well as some linear regressions using R.

Part 2 Importing and Graphing Data In R