Eight to Late

Sensemaking and Analytics for Organizations

Archive for the ‘Statistics’ Category

A gentle introduction to cluster analysis using R

with 21 comments


Welcome to the second part of my introductory series on text analysis using R (the first article can be accessed here).  My aim in the present piece is to provide a  practical introduction to cluster analysis. I’ll begin with some background before moving on to the nuts and bolts of clustering. We have a fair bit to cover, so let’s get right to it.

A common problem when analysing large collections of documents is to categorize them in some meaningful way. This is easy enough if one has a predefined classification scheme that is known to fit the collection (and if the collection is small enough to be browsed manually). One can then simply scan the documents, looking for keywords appropriate to each category and classify the documents based on the results. More often than not, however, such a classification scheme is not available and the collection too large. One then needs to use algorithms that can classify documents automatically based on their structure and content.

The present post is a practical introduction to a couple of automatic text categorization techniques, often referred to as clustering algorithms.  As the Wikipedia article on clustering tells us:

Cluster analysis or clustering is the task of grouping a set of objects in such a way that objects in the same group (called a cluster) are more similar (in some sense or another) to each other than to those in other groups (clusters).

As one might guess from the above, the results of clustering depend rather critically on the method one uses to group objects. Again, quoting from the Wikipedia piece:

Cluster analysis itself is not one specific algorithm, but the general task to be solved. It can be achieved by various algorithms that differ significantly in their notion of what constitutes a cluster and how to efficiently find them. Popular notions of clusters include groups with small distances [Note: we’ll use distance-based methods] among the cluster members, dense areas of the data space, intervals or particular statistical distributions [i.e. distributions of words within documents and the entire collection].

…and a bit later:

…the notion of a “cluster” cannot be precisely defined, which is one of the reasons why there are so many clustering algorithms. There is a common denominator: a group of data objects. However, different researchers employ different cluster models, and for each of these cluster models again different algorithms can be given. The notion of a cluster, as found by different algorithms, varies significantly in its properties. Understanding these “cluster models” is key to understanding the differences between the various algorithms.

An upshot of the above is that it is not always straightforward to interpret the output of clustering algorithms. Indeed, we will see this in the example discussed below.

With that said for an introduction, let’s move on to the nut and bolts of clustering.

Preprocessing the corpus

In this section I cover the steps required to create the R objects necessary in order to do clustering. It goes over territory that I’ve covered in detail in the first article in this series – albeit with a few tweaks, so you may want to skim through even if you’ve read my previous piece.

To begin with I’ll assume you have R and RStudio (a free development environment for R) installed on your computer and are familiar with the basic functionality in the text mining ™ package.  If you need help with this, please look at the instructions in my previous article on text mining.

As in the first part of this series,  I will use 30 posts from my blog as the example collection (or corpus, in text mining-speak). The corpus can be downloaded here. For completeness, I will run through the entire sequence of steps – right from loading the corpus into R, to running the two clustering algorithms.

Ready? Let’s go…

The first step is to fire up RStudio and navigate to the directory in which you have unpacked the example corpus. Once this is done, load the text mining package, tm.  Here’s the relevant code (Note: a complete listing of the code in this article can be accessed here):

[1] “C:/Users/Kailash/Documents”

#set working directory – fix path as needed!
#load tm library

Loading required package: NLP

Note: R commands are in blue, output in black or red; lines that start with # are comments.

If you get an error here, you probably need to download and install the tm package. You can do this in RStudio by going to Tools > Install Packages and entering “tm”. When installing a new package, R automatically checks for and installs any dependent packages.

The next step is to load the collection of documents into an object that can be manipulated by functions in the tm package.

#Create Corpus
docs <- Corpus(DirSource(“C:/Users/Kailash/Documents/TextMining”))
#inspect a particular document

The next step is to clean up the corpus. This includes things such as transforming to a consistent case, removing non-standard symbols & punctuation, and removing numbers (assuming that numbers do not contain useful information, which is the case here):

#Transform to lower case
docs <- tm_map(docs,content_transformer(tolower))
#remove potentiallyy problematic symbols
toSpace <- content_transformer(function(x, pattern) { return (gsub(pattern, ” “, x))})
docs <- tm_map(docs, toSpace, “-“)
docs <- tm_map(docs, toSpace, “:”)
docs <- tm_map(docs, toSpace, “‘”)
docs <- tm_map(docs, toSpace, “•”)
docs <- tm_map(docs, toSpace, “•    “)
docs <- tm_map(docs, toSpace, ” -“)
docs <- tm_map(docs, toSpace, ““”)
docs <- tm_map(docs, toSpace, “””)
#remove punctuation
docs <- tm_map(docs, removePunctuation)
#Strip digits
docs <- tm_map(docs, removeNumbers)

Note: please see my previous article for more on content_transformer and the toSpace function defined above.

Next we remove stopwords – common words (like “a” “and” “the”, for example) and eliminate extraneous whitespaces.

#remove stopwords
docs <- tm_map(docs, removeWords, stopwords(“english”))
#remove whitespace
docs <- tm_map(docs, stripWhitespace)

flexibility eye beholder action increase organisational flexibility say redeploying employees likely seen affected move constrains individual flexibility dual meaning characteristic many organizational platitudes excellence synergy andgovernance interesting exercise analyse platitudes expose difference espoused actual meanings sign wishing many hours platitude deconstructing fun

At this point it is critical to inspect the corpus because  stopword removal in tm can be flaky. Yes, this is annoying but not a showstopper because one can remove problematic words manually once one has identified them – more about this in a minute.

Next, we stem the document – i.e. truncate words to their base form. For example, “education”, “educate” and “educative” are stemmed to “educat.”:

docs <- tm_map(docs,stemDocument)

Stemming works well enough, but there are some fixes that need to be done due to my inconsistent use of British/Aussie and US English. Also, we’ll take this opportunity to fix up some concatenations like “andgovernance” (see paragraph printed out above). Here’s the code:

docs <- tm_map(docs, content_transformer(gsub),pattern = “organiz”, replacement = “organ”)
docs <- tm_map(docs, content_transformer(gsub), pattern = “organis”, replacement = “organ”)
docs <- tm_map(docs, content_transformer(gsub), pattern = “andgovern”, replacement = “govern”)
docs <- tm_map(docs, content_transformer(gsub), pattern = “inenterpris”, replacement = “enterpris”)
docs <- tm_map(docs, content_transformer(gsub), pattern = “team-“, replacement = “team”)

The next step is to remove the stopwords that were missed by R. The best way to do this  for a small corpus is to go through it and compile a list of words to be eliminated. One can then create a custom vector containing words to be removed and use the removeWords transformation to do the needful. Here is the code (Note:  + indicates a continuation of a statement from the previous line):

myStopwords <- c(“can”, “say”,”one”,”way”,”use”,
+                  “also”,”howev”,”tell”,”will”,
+                  “much”,”need”,”take”,”tend”,”even”,
+                  “like”,”particular”,”rather”,”said”,
+                  “get”,”well”,”make”,”ask”,”come”,”end”,
+                  “first”,”two”,”help”,”often”,”may”,
+                  “might”,”see”,”someth”,”thing”,”point”,
+                  “post”,”look”,”right”,”now”,”think”,”’ve “,
+                  “’re “)
#remove custom stopwords
docs <- tm_map(docs, removeWords, myStopwords)

Again, it is a good idea to check that the offending words have really been eliminated.

The final preprocessing step is to create a document-term matrix (DTM) – a matrix that lists all occurrences of words in the corpus.  In a DTM, documents are represented by rows and the terms (or words) by columns.  If a word occurs in a particular document n times, then the matrix entry for corresponding to that row and column is n, if it doesn’t occur at all, the entry is 0.

Creating a DTM is straightforward– one simply uses the built-in DocumentTermMatrix function provided by the tm package like so:

dtm <- DocumentTermMatrix(docs)
#print a summary

Non-/sparse entries: 13312/110618
Sparsity           : 89%
Maximal term length: 48
Weighting          : term frequency (tf)

This brings us to the end of the preprocessing phase. Next, I’ll briefly explain how distance-based algorithms work before going on to the actual work of clustering.

An intuitive introduction to the algorithms

As mentioned in the introduction, the basic idea behind document or text clustering is to categorise documents into groups based on likeness. Let’s take a brief look at how the algorithms work their magic.

Consider the structure of the DTM. Very briefly, it is a matrix in which the documents are represented as rows and words as columns. In our case, the corpus has 30 documents and 4131 words, so the DTM is a 30 x 4131 matrix.  Mathematically, one can think of this matrix as describing a 4131 dimensional space in which each of the words represents a coordinate axis and each document is represented as a point in this space. This is hard to visualise of course, so it may help to illustrate this via a two-document corpus with only three words in total.

Consider the following corpus:

Document A: “five plus five”

Document B: “five plus six”

These two  documents can be represented as points in a 3 dimensional space that has the words “five” “plus” and “six” as the three coordinate axes (see figure 1).

Figure 1: Documents A and B as points in a 3-word space

Figure 1: Documents A and B as points in a 3-word space

Now, if each of the documents can be thought of as a point in a space, it is easy enough to take the next logical step which is to define the notion of a distance between two points (i.e. two documents). In figure 1 the distance between A and B  (which I denote as D(A,B))is the length of the line connecting the two points, which is simply, the sum of the squares of the differences between the coordinates of the two points representing the documents.

D(A,B) = \sqrt{(2-1)^2 + (1-1)^2+(0-1)^2} = \sqrt 2

Generalising the above to the 4131 dimensional space at hand, the distance between two documents (let’s call them X and Y) have coordinates (word frequencies)  (x_1,x_2,...x_{4131}) and (y_1,y_2,...y_{4131}), then one can define the straight line distance (also called Euclidean distance)  D(X,Y) between them as:

D(X,Y) = \sqrt{(x_1 - y_1)^2+(x_2 - y_2)^2+...+(x_{4131} - y_{4131})^2}

It should be noted that the Euclidean distance that I have described is above is not the only possible way to define distance mathematically. There are many others but it would take me too far afield to discuss them here – see this article for more  (and don’t be put off by the term metric,  a metric  in this context is merely a distance)

What’s important here is the idea that one can define a numerical distance between documents. Once this is grasped, it is easy to understand the basic idea behind how (some) clustering algorithms work – they group documents based on distance-related criteria.  To be sure, this explanation is simplistic and glosses over some of the complicated details in the algorithms. Nevertheless it is a reasonable, approximate explanation for what goes on under the hood. I hope purists reading this will agree!

Finally, for completeness I should mention that there are many clustering algorithms out there, and not all of them are distance-based.

Hierarchical clustering

The first algorithm we’ll look at is hierarchical clustering. As the Wikipedia article on the topic tells us, strategies for hierarchical clustering fall into two types:

Agglomerative: where we start out with each document in its own cluster. The algorithm  iteratively merges documents or clusters that are closest to each other until the entire corpus forms a single cluster. Each merge happens at a different (increasing) distance.

Divisive:  where we start out with the entire set of documents in a single cluster. At each step  the algorithm splits the cluster recursively until each document is in its own cluster. This is basically the inverse of an agglomerative strategy.

The algorithm we’ll use is hclust which does agglomerative hierarchical clustering. Here’s a simplified description of how it works:

  1. Assign each document to its own (single member) cluster
  2. Find the pair of clusters that are closest to each other and merge them. So you now have one cluster less than before.
  3. Compute distances between the new cluster and each of the old clusters.
  4. Repeat steps 2 and 3 until you have a single cluster containing all documents.

We’ll need to do a few things before running the algorithm. Firstly, we need to convert the DTM into a standard matrix which can be used by dist, the distance computation function in R (the DTM is not stored as a standard matrix). We’ll also shorten the document names so that they display nicely in the graph that we will use to display results of hclust (the names I have given the documents are just way too long). Here’s the relevant code:

#convert dtm to matrix
m <- as.matrix(dtm)
#write as csv file (optional)
#shorten rownames for display purposes
rownames(m) <- paste(substring(rownames(m),1,3),rep(“..”,nrow(m)),
+                      substring(rownames(m), nchar(rownames(m))-12,nchar(rownames(m))-4))
#compute distance between document vectors
d <- dist(m)

Next we run hclust. The algorithm offers several options check out the documentation for details. I use a popular option called Ward’s method – there are others, and I suggest you experiment with them  as each of them gives slightly different results making interpretation somewhat tricky (did I mention that clustering is as much an art as a science??). Finally, we visualise the results in a dendogram (see Figure 2 below).

#run hierarchical clustering using Ward’s method
groups <- hclust(d,method=”ward.D”)
#plot dendogram, use hang to ensure that labels fall below tree
plot(groups, hang=-1)
Figure 2: Dendogram from hierarchical clustering of corpus

Figure 2: Dendogram from hierarchical clustering of corpus

A few words on interpreting dendrograms for hierarchical clusters: as you work your way down the tree in figure 2, each branch point you encounter is the distance at which a cluster merge occurred. Clearly, the most well-defined clusters are those that have the largest separation; many closely spaced branch points indicate a lack of dissimilarity (i.e. distance, in this case) between clusters. Based on this, the figure reveals that there are 2 well-defined clusters – the first one consisting of the three documents at the right end of the cluster and the second containing all other documents. We can display the clusters on the graph using the rect.hclust function like so:

#cut into 2 subtrees – try 3 and 5

The result is shown in the figure below.

Figure 3: 2 cluster solution

Figure 3: 2 cluster grouping

The figures 4 and 5 below show the grouping for 3,  and 5 clusters.

Figure 4: 3 cluster solution

Figure 4: 3 cluster grouping

Figure 5: 5 cluster solution

Figure 5: 5 cluster grouping

I’ll make just one point here: the 2 cluster grouping seems the most robust one as it happens at large distance, and is cleanly separated (distance-wise) from the 3 and 5 cluster grouping. That said, I’ll leave you to explore the ins and outs of hclust on your own and move on to our next algorithm.

K means clustering

In hierarchical clustering we did not specify the number of clusters upfront. These were determined by looking at the dendogram after the algorithm had done its work.  In contrast, our next algorithm – K means –   requires us to define the number of clusters upfront (this number being the “k” in the name). The algorithm then generates k document clusters in a way that ensures the within-cluster distances from each cluster member to the centroid (or geometric mean) of the cluster is minimised.

Here’s a simplified description of the algorithm:

  1. Assign the documents randomly to k bins
  2. Compute the location of the centroid of each bin.
  3. Compute the distance between each document and each centroid
  4. Assign each document to the bin corresponding to the centroid closest to it.
  5. Stop if no document is moved to a new bin, else go to step 2.

An important limitation of the k means method is that the solution found by the algorithm corresponds to a local rather than global minimum (this figure from Wikipedia explains the difference between the two in a nice succinct way). As a consequence it is important to run the algorithm a number of times (each time with a different starting configuration) and then select the result that gives the overall lowest sum of within-cluster distances for all documents.  A simple check that a solution is robust is to run the algorithm for an increasing number of initial configurations until the result does not change significantly. That said, this procedure does not guarantee a globally optimal solution.

I reckon that’s enough said about the algorithm, let’s get on with it using it. The relevant function, as you might well have guessed is kmeans. As always, I urge you to check the documentation to understand the available options. We’ll use the default options for all parameters excepting nstart which we set to 100. We also plot the result using the clusplot function from the cluster library (which you may need to install. Reminder you can install packages via the Tools>Install Packages menu in RStudio)

#k means algorithm, 2 clusters, 100 starting configurations
kfit <- kmeans(d, 2, nstart=100)
#plot – need library cluster
clusplot(m, kfit$cluster, color=T, shade=T, labels=2, lines=0)

The plot is shown in Figure 6.

Figure 6: principal component plot (k=2)

Figure 6: principal component plot (k=2)

The cluster plot shown in the figure above needs a bit of explanation. As mentioned earlier, the clustering algorithms work in a mathematical space whose dimensionality equals the number of words in the corpus (4131 in our case). Clearly, this is impossible to visualize.  To handle this, mathematicians have invented a dimensionality reduction technique called Principal Component Analysis which reduces the number of dimensions to 2 (in this case) in such a way that the reduced dimensions capture as much of the variability between the clusters as possible (and hence the comment, “these two components explain 69.42% of the point variability” at the bottom of the plot in figure 6)

(Aside  Yes I realize the figures are hard to read because of the overly long names, I leave it to you to fix that. No excuses, you know how…:-))

Running the algorithm and plotting the results for k=3 and 5 yields the figures below.

Figure 7: Principal component plot (k=3)

Figure 7: Principal component plot (k=3)

Figure 8: Principal component plot (k=5)

Figure 8: Principal component plot (k=5)

Choosing k

Recall that the k means algorithm requires us to specify k upfront. A natural question then is: what is the best choice of k? In truth there is no one-size-fits-all answer to this question, but there are some heuristics that might sometimes help guide the choice. For completeness I’ll describe one below even though it is not much help in our clustering problem.

In my simplified description of the k means algorithm I mentioned that the technique attempts to minimise the sum of the distances between the points in a cluster and the cluster’s centroid. Actually, the quantity that is minimised is the total of the within-cluster sum of squares (WSS) between each point and the mean. Intuitively one might expect this quantity to be maximum when k=1 and then decrease as k increases, sharply at first and then less sharply as k reaches its optimal value.

The problem with this reasoning is that it often happens that the within cluster sum of squares never shows a slowing down in decrease of the summed WSS. Unfortunately this is exactly what happens in the case at hand.

I reckon a picture might help make the above clearer. Below is the R code to draw a plot of summed WSS as a function of k for k=2 all the way to 29 (1-total number of documents):

#kmeans – determine the optimum number of clusters (elbow method)
#look for “elbow” in plot of summed intra-cluster distances (withinss) as fn of k
wss <- 2:29
for (i in 2:29) wss[i] <- sum(kmeans(d,centers=i,nstart=25)$withinss)
plot(2:29, wss[2:29], type=”b”, xlab=”Number of Clusters”,ylab=”Within groups sum of squares”)

…and the figure below shows the resulting plot.

Figure 10: WSS as a function of k (“elbow plot”)

The plot clearly shows that there is no k for which the summed WSS flattens out (no distinct “elbow”).  As a result this method does not help. Fortunately, in this case  one can get a sensible answer using common sense rather than computation:  a choice of 2 clusters seems optimal because both algorithms yield exactly the same clusters and show the clearest cluster separation at this point (review the dendogram and cluster plots for k=2).

The meaning of it all

Now I must acknowledge an elephant in the room that I have steadfastly ignored thus far. The odds are good that you’ve seen it already….

It is this: what topics or themes do the (two) clusters correspond to?

Unfortunately this question does not have a straightforward answer. Although the algorithms suggest a 2-cluster grouping, they are silent on the topics or themes related to these.   Moreover,  as you will see if you experiment, the results of clustering depend on:

  • The criteria for the construction of the DTM  (see the documentation for DocumentTermMatrix for options).
  • The clustering algorithm itself.

Indeed, insofar as clustering is concerned, subject matter and corpus knowledge is the best way to figure out cluster themes. This serves to reinforce (yet again!) that clustering is as much an art as it is a science.

In the case at hand, article length seems to be an important differentiator between the 2 clusters found by both algorithms. The three articles in the smaller cluster are in the top 4 longest pieces in the corpus.  Additionally, the three pieces are related to sensemaking and dialogue mapping. There are probably other factors as well, but none that stand out as being significant. I should mention, however, that the fact that article length seems to play a significant role here suggests that it may be worth checking out the effect of scaling distances by word counts or using other measures such a cosine similarity – but that’s a topic for another post! (Note added on Dec 3 2015: check out my article on visualizing relationships between documents using network graphs for a detailed discussion on cosine similarity)

The take home lesson is that  is that the results of clustering are often hard to interpret. This should not be surprising – the algorithms cannot interpret meaning, they simply chug through a mathematical optimisation problem. The onus is on the analyst to figure out what it means…or if it means anything at all.


This brings us to the end of a long ramble through clustering.  We’ve explored the two most common methods:  hierarchical and k means clustering (there are many others available in R, and I urge you to explore them). Apart from providing the detailed steps to do clustering, I have attempted to provide an intuitive explanation of how the algorithms work.  I hope I have succeeded in doing so. As always your feedback would be very welcome.

Finally, I’d like to reiterate an important point:  the results of our clustering exercise do not have a straightforward interpretation, and this is often the case in cluster analysis. Fortunately I can close on an optimistic note. There are other text mining techniques that do a better job in grouping documents based on topics and themes rather than word frequencies alone.   I’ll discuss this in the next article in this series.  Until then, I wish you many enjoyable hours exploring the ins and outs of clustering.

Note added on September 29th 2015:

If you liked this article, you might want to check out its sequel – an introduction to topic modeling.

Written by K

July 22, 2015 at 8:53 pm

A gentle introduction to text mining using R

with 43 comments


This article is based on my exploration of the basic text mining capabilities of  R, the open source statistical software. It is intended primarily as a tutorial  for novices in text mining as well as R. However, unlike conventional tutorials,  I spend a good bit of time setting the context by describing the problem that led me to text mining and thence  to R. I also talk about the limitations of  the techniques I describe,  and point out directions for further exploration for those who are interested. Indeed, I’ll likely explore some of these myself in future articles.

If you have no time and /or wish to cut to the chase,  please go straight to the section entitled, Preliminaries – installing R and RStudio. If you have already installed R and have worked with it, you may want to stop reading as I  doubt there’s anything I can tell you that you don’t already know 🙂

A couple of warnings are in order before we proceed. R and the text mining options we explore below are open source software. Version differences in open source can be significant and are not always documented in a way that corporate IT types are used to. Indeed, I was tripped up by such differences in an earlier version of this article (now revised). So, just for the record, the examples below were run on version 3.2.0 of R and version 0.6-1 of the tm (text mining) package for R. A second point follows from this: as is evident from its version number, the tm package is still in the early stages of its evolution. As a result – and we will see this below – things do not always work as advertised. So assume nothing, and inspect the results in detail at every step. Be warned that I do not always do this below, as my aim is introduction rather than accuracy.

Background and motivation

Traditional data analysis is based on the relational model in which data is stored in tables. Within tables, data is stored in rows – each row representing a  single record of an entity of interest (such as a customer or an account). The columns represent attributes of the entity. For example, the customer table might consist of columns such as name, street address, city, postcode, telephone number .  Typically these  are defined upfront, when the data model is created. It is possible to add columns after the fact, but this tends to be messy because one also has to update existing rows with information pertaining to the added attribute.

As long as one asks for information that is based  only on existing attributes – an example being,  “give me a list of customers based  in Sydney” –  a database analyst can use Structured Query Language (the defacto language of relational databases ) to  get an answer.  A problem arises, however, if one asks for information that is based on attributes that are not included in the database. An example  in the above case would be: “give me a list of  customers who have made a complaint in the last twelve months.”

As a result of the above, many data modelers will include  a “catch-all”  free text column that can be used to capture additional information in an ad-hoc way. As one might imagine, this column will often end up containing several lines, or even paragraphs of text that are near impossible to analyse with the tools available in relational databases.

(Note: for completeness I should add that most database vendors have incorporated text mining capabilities into their products. Indeed, many  of them now include R…which is another good reason to learn it.)

My story

Over the last few months, when time permits, I’ve been doing an in-depth exploration of  the data captured by my organisation’s IT service management tool.  Such tools capture all support tickets that are logged, and track their progress until they are closed.  As it turns out,  there are a number of cases where calls are logged against categories that are too broad to be useful – the infamous catch-all category called “Unknown.”  In such cases, much of the important information is captured in a free text column, which is difficult to analyse unless one knows what one is looking for. The problem I was grappling with was to identify patterns and hence define sub-categories that would enable support staff to categorise these calls meaningfully.

One way to do this  is  to guess what the sub-categories might be…and one can sometimes make pretty good guesses if one knows the data well enough.  In general, however, guessing is a terrible strategy because one does not know what one does not know. The only sensible way to extract subcategories is to  analyse  the content of the free text column systematically. This is a classic text mining problem.

Now, I knew a bit about the theory of text mining, but had little practical experience with it. So the logical place for me to  start was to look for a suitable text mining tool. Our vendor (who shall remain unnamed) has a “Rolls-Royce” statistical tool that has a good text mining add-on. We don’t have licenses for the tool, but the vendor was willing to give us a trial license for a few months…with the understanding that this was on an intent-to-purchase basis.

I therefore started looking at open source options. While doing so, I stumbled on an interesting paper by Ingo Feinerer that describes a text mining framework for the R environment. Now, I knew about R, and was vaguely aware that it offered text mining capabilities, but I’d not looked into the details.  Anyway, I started reading the paper…and kept going until I finished.

As I read, I realised that this could be the answer to my problems. Even better, it would not require me trade in assorted limbs for a license.

I decided to give it a go.

Preliminaries – installing R and RStudio

R can be downloaded from the R Project website. There is a Windows version available, which installed painlessly on my laptop. Commonly encountered installation issues are answered in the (very helpful)  R for Windows FAQ.

RStudio is an integrated development environment (IDE) for R. There is a commercial version of the product, but there is also a free open source version. In what follows, I’ve used the free version. Like R, RStudio installs painlessly and also detects your R installation.

RStudio has  the following panels:

  • A script editor in which you can create R scripts (top left). You can also open a new script editor window by going to File > New File > RScript.
  • The console where you can execute R commands/scripts (bottom left)
  • Environment and history (top right)
  • Files in the current working directory, installed R packages, plots and a help display screen (bottom right).

Check out this short video for a quick introduction to RStudio.

You can access help anytime (within both R and RStudio) by typing a question mark before a command. Exercise: try this by typing ?getwd() and ?setwd() in the console.

I should reiterate that the installation process for both products was seriously simple…and seriously impressive.  “Rolls-Royce” business intelligence vendors could take a lesson from  that…in addition to taking a long hard look at the ridiculous prices they charge.

There is another small step before we move on to the fun stuff.   Text mining  and certain plotting packages are not installed by default so one has to install them manually The relevant packages are:

  1. tm – the text mining package (see documentation). Also check out this excellent introductory article on tm.
  2. SnowballC – required for stemming (explained below).
  3. ggplot2 – plotting capabilities (see documentation)
  4. wordcloud – which is self-explanatory (see documentation) .

(Warning for Windows users: R is case-sensitive so Wordcloud != wordcloud)

The simplest way to install packages is to use RStudio’s built in capabilities (go to Tools > Install Packages in the menu). If you’re working on Windows 7 or 8, you might run into a permissions issue when installing packages. If you do, you might find this advice from the R for Windows FAQ helpful.

Preliminaries – The example dataset

The data I had from our service management tool isn’t  the best dataset to learn with as it is quite messy. But then, I  have a reasonable data source in my virtual backyard:  this blog. To this end, I converted all posts I’ve written since Dec 2013 into plain text form (30 posts in all). You can download the zip file of these here .

I suggest you create a new folder called – called, say, TextMining – and unzip the files in that folder.

That done, we’re good to start…

Preliminaries – Basic Navigation

A few things to note before we proceed:

  • In what follows, I enter the commands directly in the console. However,  here’s a little RStudio tip that you may want to consider: you can enter an R command or code fragment in the script editor and then hit Ctrl-Enter  (i.e. hit the Enter key while holding down the Control key) to copy the line to the console.  This will enable you to save the script as you go along.
  • In the code snippets below, the functions / commands to be typed in the R console are in blue font.   The output is in black. I will also denote references to  functions / commands in the body of the article by italicising them as in “setwd()”.  Be aware that I’ve omitted the command prompt “>” in the code snippets below!
  • It is best not to cut-n-paste commands directly from the article as quotes are sometimes not rendered correctly. A text file of all the code in this article is available here.

The > prompt in the RStudio console indicates that R is ready to process commands.

To see the current working directory type in getwd() and hit return. You’ll see something like:

[1] “C:/Users/Documents”


The exact output will of course depend on your working directory.  Note the forward slashes in the path. This is because of R’s Unix heritage (backslash is an escape character in R.). So, here’s how would change the working directory to C:\Users: 


You can now use getwd()to check that setwd() has done what it should.



I won’t say much more here about R  as I want to get on with the main business of the article.  Check out this very short introduction to R for a quick crash course.

Loading data into R

Start RStudio and open the TextMining project you created earlier.

The next step is to load the tm package as this is not loaded by default.  This is done using the library() function like so:

Loading required package: NLP


Dependent packages are loaded automatically – in this case the dependency is on the NLP (natural language processing) package.

Next, we need to create a collection of documents (technically referred to as a Corpus) in the R environment. This basically involves loading the files created in the TextMining folder into a Corpus object. The tm package provides the Corpus() function to do this. There are several ways to  create a Corpus (check out the online help using ? as explained earlier). In a nutshell, the  Corpus() function can read from various sources including a directory. That’s the option we’ll use:

#Create Corpus
docs <- Corpus(DirSource(“C:/Users/Kailash/Documents/TextMining”))


At the risk of stating the obvious, you will need to tailor this path as appropriate.

A couple of things to note in the above. Any line that starts with a # is a comment, and the “<-“ tells R to assign the result of the command on the right hand side to the variable on the left hand side. In this case the Corpus object created is stored in a variable called docs.  One can also use the equals sign (=)  for assignment if one wants to.

Type in docs to see some information about the newly created corpus:

Metadata: corpus specific: 0, document level (indexed): 0
Content: documents: 30


The summary() function gives more details, including a complete listing of files…but it isn’t particularly enlightening.  Instead, we’ll examine a particular document in the corpus.

#inspect a particular document
…output not shown…

Which prints the entire content of 30th document in the corpus to the console.


Data cleansing, though tedious, is perhaps the most important step in text analysis.   As we will see, dirty data can play havoc with the results.  Furthermore, as we will also see, data cleaning is invariably an iterative process as there are always problems that are overlooked the first time around.

The tm package offers a number of transformations that ease the tedium of cleaning data. To see the available transformations  type getTransformations() at the R prompt:

> getTransformations()
[1] “removeNumbers” “removePunctuation” “removeWords” “stemDocument” “stripWhitespace”


Most of these are self-explanatory. I’ll explain those that aren’t as we go along.

There are a few preliminary clean-up steps we need to do before we use these powerful transformations. If you inspect some documents in the corpus (and you know how to do that now), you will notice that I have some quirks in my writing. For example, I often use colons and hyphens without spaces between the words separated by them. Using the removePunctuation transform  without fixing this will cause the two words on either side of the symbols  to be combined. Clearly, we need to fix this prior to using the transformations.

To fix the above, one has to create a custom transformation. The tm package provides the ability to do this via the content_transformer function. This function takes a function as input, the input function should specify what transformation needs to be done. In this case, the input function would be one that replaces all instances of a character by spaces. As it turns out the gsub() function does just that.

Here is the R code to build the content transformer, which  we will call toSpace:

#create the toSpace content transformer
toSpace <- content_transformer(function(x, pattern) {return (gsub(pattern, ” “, x))})

Now we can use  this content transformer to eliminate colons and hypens like so:

docs <- tm_map(docs, toSpace, “-“)
docs <- tm_map(docs, toSpace, “:”)


Inspect random sections f corpus to check that the result is what you intend (use writeLines as shown earlier). To reiterate something I mentioned in the preamble, it is good practice to inspect the a subset of the corpus after each transformation.

If it all looks good, we can now apply the removePunctuation transformation. This is done as follows:

#Remove punctuation – replace punctuation marks with ” “
docs <- tm_map(docs, removePunctuation)


Inspecting the corpus reveals that several  “non-standard” punctuation marks have not been removed. These include the single curly quote marks and a space-hyphen combination. These can be removed using our custom content transformer, toSpace. Note that you might want to copy-n-paste these symbols directly from the relevant text file to ensure that they are accurately represented in toSpace.

docs <- tm_map(docs, toSpace, “’”)
docs <- tm_map(docs, toSpace, “‘”)
docs <- tm_map(docs, toSpace, ” -“)


Inspect the corpus again to ensure that the offenders have been eliminated. This is also a good time to check for any other special symbols that may need to be removed manually.

If all is well, you can move  to the next step which is  to:

  • Convert the corpus to lower case
  • Remove all numbers.

Since R is case sensitive, “Text” is not equal to “text” – and hence the rationale for converting to a standard case.  However, although there is a tolower transformation, it is not a part of the standard tm transformations (see the output of getTransformations() in the previous section). For this reason, we have to convert tolower into a transformation that can handle a corpus object properly. This is done with the help of our new friend, content_transformer.

Here’s the relevant code:

#Transform to lower case (need to wrap in content_transformer)
docs <- tm_map(docs,content_transformer(tolower))

Text analysts are typically not interested in numbers since these do not usually contribute to the meaning of the text. However, this may not always be so. For example, it is definitely not the case if one is interested in getting a count of the number of times a particular year appears in a corpus. This does not need to be wrapped in content_transformer as it is a standard transformation in tm.

#Strip digits (std transformation, so no need for content_transformer)
docs <- tm_map(docs, removeNumbers)

Once again, be sure to inspect the corpus before proceeding.

The next step is to remove common words  from the text. These  include words such as articles (a, an, the), conjunctions (and, or but etc.), common verbs (is), qualifiers (yet, however etc) . The tm package includes  a standard list of such stop words as they are referred to. We remove stop words using the standard removeWords transformation like so:

#remove stopwords using the standard list in tm
docs <- tm_map(docs, removeWords, stopwords(“english”))


Finally, we remove all extraneous whitespaces using the stripWhitespace transformation:

#Strip whitespace (cosmetic?)
docs <- tm_map(docs, stripWhitespace)



Typically a large corpus will contain  many words that have a common root – for example: offer, offered and offering.  Stemming is the process of reducing such related words to their common root, which in this case would be the word offer.

Simple stemming algorithms (such as the one in tm) are relatively crude: they work by chopping off the ends of words. This can cause problems: for example, the words mate and mating might be reduced to mat instead of mate.  That said, the overall benefit gained from stemming more than makes up for the downside of such special cases.

To see what stemming does, let’s take a look at the  last few lines  of the corpus before and after stemming.  Here’s what the last bit looks  like prior to stemming (note that this may differ for you, depending on the ordering of the corpus source files in your directory):

flexibility eye beholder action increase organisational flexibility say redeploying employees likely seen affected move constrains individual flexibility dual meaning characteristic many organizational platitudes excellence synergy andgovernance interesting exercise analyse platitudes expose difference espoused actual meanings sign wishing many hours platitude deconstructing fun 


Now let’s stem the corpus and reinspect it.

#load library
#Stem document
docs <- tm_map(docs,stemDocument)
flexibl eye behold action increas organis flexibl say redeploy employe like seen affect move constrain individu flexibl dual mean characterist mani organiz platitud excel synergi andgovern interest exercis analys platitud expos differ espous actual mean sign wish mani hour platitud deconstruct fun


A careful comparison of the two paragraphs reveals the benefits and tradeoff of this relatively crude process.

There is a more sophisticated procedure called lemmatization that takes grammatical context into account. Among other things, determining the lemma of a word requires a knowledge of its part of speech (POS) – i.e. whether it is a noun, adjective etc. There are POS taggers that automate the process of tagging terms with their parts of speech. Although POS taggers are available for R (see this one, for example), I will not go into this topic here as it would make a long post even longer.

On another important note, the output of the corpus also shows up a problem or two. First, organiz and organis are actually variants of the same stem organ. Clearly, they should be merged. Second, the word andgovern should be separated out into and and govern (this is an error in the original text).  These (and other errors of their ilk) can and should be fixed up before proceeding.  This is easily done using gsub() wrapped in content_transformer. Here is the code to  clean up these and a few other issues  that I found:

docs <- tm_map(docs, content_transformer(gsub), pattern = “organiz”, replacement = “organ”)
docs <- tm_map(docs, content_transformer(gsub), pattern = “organis”, replacement = “organ”)
docs <- tm_map(docs, content_transformer(gsub), pattern = “andgovern”, replacement = “govern”)
docs <- tm_map(docs, content_transformer(gsub), pattern = “inenterpris”, replacement = “enterpris”)
docs <- tm_map(docs, content_transformer(gsub), pattern = “team-“, replacement = “team”)


Note that I have removed the stop words and and in in the 3rd and 4th transforms above.

There are definitely other errors that need to be cleaned up, but I’ll leave these for you to detect and remove.

The document term matrix

The next step in the process is the creation of the document term matrix  (DTM)– a matrix that lists all occurrences of words in the corpus, by document. In the DTM, the documents are represented by rows and the terms (or words) by columns.  If a word occurs in a particular document, then the matrix entry for corresponding to that row and column is 1, else it is 0 (multiple occurrences within a document are recorded – that is, if a word occurs twice in a document, it is recorded as “2” in the relevant matrix entry).

A simple example might serve to explain the structure of the TDM more clearly. Assume we have a simple corpus consisting of two documents, Doc1 and Doc2, with the following content:

Doc1: bananas are yellow

Doc2: bananas are good

The DTM for this corpus would look like:

bananas are yellow good
Doc1 1 1 1 0
Doc2 1 1 0 1


Clearly there is nothing special about rows and columns – we could just as easily transpose them. If we did so, we’d get a term document matrix (TDM) in which the terms are rows and documents columns. One can work with either a DTM or TDM. I’ll use the DTM in what follows.

There are a couple of general points worth making before we proceed. Firstly, DTMs (or TDMs) can be huge – the dimension of the matrix would be number of document  x the number of words in the corpus.  Secondly, it is clear that the large majority of words will appear only in a few documents. As a result a DTM is invariably sparse – that is, a large number of its entries are 0.

The business of creating a DTM (or TDM) in R is as simple as:

dtm <- DocumentTermMatrix(docs)


This creates a term document matrix from the corpus and stores the result in the variable dtm. One can get summary information on the matrix by typing the variable name in the console and hitting return:

<<DocumentTermMatrix (documents: 30, terms: 4209)>>
Non-/sparse entries: 14252/112018
Sparsity : 89%
Maximal term length: 48
Weighting : term frequency (tf)


This is a 30 x 4209 dimension matrix in which 89% of the rows are zero.

One can inspect the DTM, and you might want to do so for fun. However, it isn’t particularly illuminating because of the sheer volume of information that will flash up on the console. To limit the information displayed, one can inspect a small section of it like so:

<<DocumentTermMatrix (documents: 2, terms: 6)>>
Non-/sparse entries: 0/12
Sparsity : 100%
Maximal term length: 8
Weighting : term frequency (tf)
Docs                               creation creativ credibl credit crimin crinkl
BeyondEntitiesAndRelationships.txt        0        0      0      0      0      0
bigdata.txt                               0        0      0      0      0      0

This command displays terms 1000 through 1005 in the first two rows of the DTM. Note that your results may differ.

Mining the corpus

Notice that in constructing the TDM, we have converted a corpus of text into a mathematical object that can be analysed using quantitative techniques of matrix algebra.  It should be no surprise, therefore, that the TDM (or DTM) is the starting point for quantitative text analysis.

For example, to get the frequency of occurrence of each word in the corpus, we simply sum over all rows to give column sums:

freq <- colSums(as.matrix(dtm))


Here we have  first converted the TDM into a mathematical matrix using the as.matrix() function. We have then summed over all rows to give us the totals for each column (term). The result is stored in the (column matrix) variable freq.

Check that the dimension of freq equals the number of terms:

#length should be total number of terms
[1] 4209


Next, we sort freq in descending order of term count:

#create sort order (descending)
ord <- order(freq,decreasing=TRUE)


Then list the most and least frequently occurring terms:

#inspect most frequently occurring terms
one organ can manag work system
314 268    244  222   202   193
#inspect least frequently occurring terms
yield yorkshir   youtub     zeno     zero    zulli
1        1        1        1        1        1


The  least frequent terms can be more interesting than one might think. This is  because terms that occur rarely are likely to be more descriptive of specific documents. Indeed, I can recall the posts in which I have referred to Yorkshire, Zeno’s Paradox and  Mr. Lou Zulli without having to go back to the corpus, but I’d have a hard time enumerating the posts in which I’ve used the word system.

There are at least a couple of ways to simple ways to strike a balance between frequency and specificity. One way is to use so-called  inverse document frequencies. A simpler approach is to  eliminate words that occur in a large fraction of corpus documents.   The latter addresses another issue that is evident in the above. We deal with this now.

Words like “can” and “one”  give us no information about the subject matter of the documents in which they occur. They can therefore be eliminated without loss. Indeed, they ought to have been eliminated by the stopword removal we did earlier. However, since such words occur very frequently – virtually in all documents – we can remove them by enforcing bounds when creating the DTM, like so:

dtmr <-DocumentTermMatrix(docs, control=list(wordLengths=c(4, 20),
bounds = list(global = c(3,27))))


Here we have told R to include only those words that occur in  3 to 27 documents. We have also enforced  lower and upper limit to length of the words included (between 4 and 20 characters).

Inspecting the new DTM:

<<DocumentTermMatrix (documents: 30, terms: 1290)>>
Non-/sparse entries: 10002/28698
Sparsity : 74%
Maximal term length: 15
Weighting : term frequency (tf)

The dimension is reduced to 30 x 1290.

Let’s calculate the cumulative frequencies of words across documents and sort as before:

freqr <- colSums(as.matrix(dtmr))
#length should be total number of terms
[1] 1290
#create sort order (asc)
ordr <- order(freqr,decreasing=TRUE)
#inspect most frequently occurring terms
organ manag work system project problem
268     222  202    193     184     171
#inspect least frequently occurring terms
wait warehous welcom whiteboard wider widespread
3           3      3          3     3          3


The results make sense: the top 6 keywords are pretty good descriptors of what my blogs is about – projects, management and systems. However, not all high frequency words need be significant. What they do, is give you an idea of potential classification terms.

That done, let’s take get a list of terms that occur at least a  100 times in the entire corpus. This is easily done using the findFreqTerms() function as follows:

[1] “action” “approach” “base” “busi” “chang” “consult” “data” “decis” “design”
[10] “develop” “differ” “discuss” “enterpris” “exampl” “group” “howev” “import” “issu”
[19] “like” “make” “manag” “mani” “model” “often” “organ” “peopl” “point”
[28] “practic” “problem” “process” “project” “question” “said” “system” “thing” “think”
[37] “time” “understand” “view” “well” “will” “work”


Here I have asked findFreqTerms() to return all terms that occur more than 80 times in the entire corpus. Note, however, that the result is ordered alphabetically, not by frequency.

Now that we have the most frequently occurring terms in hand, we can check for correlations between some of these and other terms that occur in the corpus.  In this context, correlation is a quantitative measure of the co-occurrence of words in multiple documents.

The tm package provides the findAssocs() function to do this.  One needs to specify the DTM, the term of interest and the correlation limit. The latter is a number between 0 and 1 that serves as a lower bound for  the strength of correlation between the  search and result terms. For example, if the correlation limit is 1, findAssocs() will return only  those words that always co-occur with the search term. A correlation limit of 0.5 will return terms that have a search term co-occurrence of at least  50% and so on.

Here are the results of  running findAssocs() on some of the frequently occurring terms (system, project, organis) at a correlation of 60%.


inher 0.82
handl 0.68
manag 0.68
occurr 0.68
manager’ 0.66
agil       0.80
realist    0.78
increment  0.77
upfront    0.77
technolog  0.70
neither    0.69
solv       0.69
adapt      0.67
architectur 0.67
happi      0.67
movement   0.67
architect  0.66
chanc      0.65
fine       0.64
featur     0.63
design  0.78
subset  0.78
adopt   0.77
user    0.77
involv  0.71
specifi 0.71
function 0.70
intend  0.67
softwar 0.67
step    0.67
compos  0.66
intent  0.66
specif  0.66
depart  0.65
phone  0.63
frequent 0.62
today  0.62
pattern 0.61
cognit 0.60
wherea 0.60


An important point to note that the presence of a term in these list is not indicative of its frequency.  Rather it is a measure of the frequency with which the two (search and result term)  co-occur (or show up together) in documents across . Note also, that it is not an indicator of nearness or contiguity. Indeed, it cannot be because the document term matrix does not store any information on proximity of terms, it is simply a “bag of words.”

That said, one can already see that the correlations throw up interesting combinations – for example, project and manag, or enterpris and agil or architect/architecture, or system and design or adopt. These give one further insights into potential classifications.

As it turned out,  the very basic techniques listed above were enough for me to get a handle on the original problem that led me to text mining – the analysis of free text problem descriptions in my organisation’s service management tool.  What I did was to work my way through the top 50 terms and find their associations. These revealed a number of sets of keywords that occurred in multiple problem descriptions,  which was good enough for me to define some useful sub-categories.  These are currently being reviewed by the service management team. While they’re busy with that that, I’m looking into refining these further using techniques such as  cluster analysis and tokenization.   A simple case of the latter would be to look at two-word combinations in the text (technically referred to as bigrams). As one might imagine, the dimensionality of the DTM will quickly get out of hand as one considers larger multi-word combinations.

Anyway,  all that and more will topics have to wait for future  articles as this piece is much too long already. That said, there is one thing I absolutely must touch upon before signing off. Do stay, I think you’ll find  it interesting.

Basic graphics

One of the really cool things about R is its graphing capability. I’ll do just a couple of simple examples to give you a flavour of its power and cool factor. There are lots of nice examples on the Web that you can try out for yourself.

Let’s first do a simple frequency histogram. I’ll use the ggplot2 package, written by Hadley Wickham to do this. Here’s the code:

p <- ggplot(subset(wf, freqr>100), aes(term, occurrences))
p <- p + geom_bar(stat=”identity”)
p <- p + theme(axis.text.x=element_text(angle=45, hjust=1))


Figure 1 shows the result.

Fig 1: Term-occurrence histogram (freq>100)

Fig 1: Term-occurrence histogram (freq>100)

The first line creates a data frame – a list of columns of equal length. A data frame also contains the name of the columns – in this case these are term and occurrence respectively.  We then invoke ggplot(), telling it to consider plot only those terms that occur more than 100 times.  The aes option in ggplot describes plot aesthetics – in this case, we use it to specify the x and y axis labels. The stat=”identity” option in geom_bar () ensures  that the height of each bar is proportional to the data value that is mapped to the y-axis  (i.e occurrences). The last line specifies that the x-axis labels should be at a 45 degree angle and should be horizontally justified (see what happens if you leave this out). Check out the voluminous ggplot documentation for more or better yet, this quick introduction to ggplot2 by Edwin Chen.

Finally, let’s create a wordcloud for no other reason than everyone who can seems to be doing it.  The code for this is:

#setting the same seed each time ensures consistent look across clouds
#limit words by specifying min frequency
wordcloud(names(freqr),freqr, min.freq=70)


The result is shown Figure 2.


Fig 2: Wordcloud (freq>70)

Fig 2: Wordcloud (freq>70)

Here we first load the wordcloud package which is not loaded by default. Setting a seed number ensures that you get the same look each time (try running it without setting a seed). The arguments of the wordcloud() function are straightforward enough. Note that one can specify the maximum number of words to be included instead of the minimum frequency (as I have done above).  See the word cloud  documentation for more.

This word cloud also makes it clear that stop word removal has not done its job well, there are a number of words it has missed (also and however, for example). These can be removed by augmenting the built-in stop word list with a custom one. This is left as an exercise for the reader :-).

Finally, one can make the wordcloud more visually appealing by adding colour as follows:

#…add color


The result is shown Figure 3.



Fig 3: Wordcloud (freq > 70)

Fig 3: Wordcloud (freq > 70)

You may need to load the RColorBrewer package to get this to work. Check out the brewer documentation to experiment with more colouring options.

Wrapping up

This brings me to the end of this rather long  (but I hope, comprehensible) introduction to text mining R.  It should be clear that despite the length of the article, I’ve covered only the most rudimentary basics.  Nevertheless, I hope I’ve succeeded in conveying  a sense of the possibilities in the vast and rapidly-expanding discipline of text analytics.


Note added on July 22nd, 2015:

If you liked this piece, you may want to check out the sequel – a gentle introduction to cluster analysis using R.

Note added on September 29th 2015:

…and my introductory piece on topic modeling.

Note added on December 3rd 2015:

…and my article on visualizing relationships between documents.

Written by K

May 27, 2015 at 8:08 pm

On the statistical downsides of blogging

with 22 comments


The stats on the 200+ posts I’ve written since I started blogging make it pretty clear that:

  1. Much of what I write does not get much attention – i.e. it is not of interest to most readers.
  2. An interesting post – a rare occurrence in itself  – is invariably followed by a series of uninteresting ones.

In this post, I ignore the very real possibility that my work is inherently uninteresting and discuss how the above observations can be explained via concepts of probability.

Base rate of uninteresting ideas

A couple of years ago I wrote a piece entitled, Trumped by Conditionality, in which I used conditional probability to show that majority of the posts on this blog will be uninteresting despite my best efforts.  My argument was based on the following observations:

  1. There are many more uninteresting ideas than interesting ones.  In statistical terminology one would say that the base rate of uninteresting ideas is high.   This implies that if I write posts without filtering out bad ideas, I will write uninteresting posts far more frequently than interesting ones.
  2. The base rate as described above is inapplicable in real life because I do attempt to filter out the bad ideas. However, and this is the key:  my ability to distinguish between interesting and uninteresting topics is imperfect. In other words, although I can generally identify an interesting idea correctly , there is a small (but significant) chance that I will incorrectly identify an uninteresting topic as being interesting.

Now,  since uninteresting ideas vastly outnumber interesting ones and my ability to filter out uninteresting ideas is imperfect, it follows that  the majority of the topics I choose to write about will be uninteresting.   This is essentially the first point I made in the introduction.

Regression to the mean

The observation that good (i.e. interesting) posts are generally followed by a series of not so good ones is a consequence of a statistical phenomenon known as regression to the mean.  In everyday language this refers to the common observation that an extreme event is generally followed by a less extreme one.   This is simply a consequence of the fact that for many commonly encountered phenomena extreme events are much less likely to occur than events that are close to the average.

In the case at hand we are concerned with the quality of writing. Although writers might improve through practice, it is pretty clear that they cannot write brilliant posts every time they put fingers to keyboard. This is particularly true of bloggers and syndicated columnists who have to produce pieces according to a timetable – regardless of practice or talent, it is impossible to produce high quality pieces on a regular basis.

It is worth noting that people often incorrectly ascribe causal explanations to phenomena that can be explained by regression to the mean.  Daniel Kahneman and Amos Tversky describe the following example in their classic paper on decision-related cognitive biases:

…In a discussion of flight training, experienced instructors noted that praise for an exceptionally smooth landing is typically followed by a poorer landing on the next try, while harsh criticism after a rough landing is usually followed by an improvement on the next try. The instructors concluded that verbal rewards are detrimental to learning, while verbal punishments are beneficial, contrary to accepted psychological doctrine. This conclusion is unwarranted because of the presence of regression toward the mean. As in other cases of repeated examination, an improvement will usually follow a poor performance and a deterioration will usually follow an outstanding performance, even if the instructor does not respond to the trainee’s achievement on the first attempt

So, although I cannot avoid the disappointment that follows the high of writing a well-received post, I can take (perhaps, false) comfort in the possibility that I’m a victim of statistics.

In closing

Finally, l would be remiss if I did not consider an explanation which, though unpleasant, may well be true: there is the distinct possibility that everything I write about is uninteresting. Needless to say, I reckon the explanations (rationalisations?) offered above are far more likely to be correct 🙂

Written by K

June 1, 2012 at 6:12 am

Posted in Probability, Statistics, Writing

Tagged with

On the accuracy of group estimates

with 19 comments


The essential idea behind group estimation is that an estimate made by a group is likely to be more accurate than one made by an individual in the group. This notion is the basis for the Delphi method and its variants. In this post, I use arguments involving probabilities to gain some insight into the conditions under which group estimates are more accurate than individual ones.

An insight from conditional probability

Let’s begin with a simple group estimation scenario.

Assume we have two individuals of similar skill who have been asked to provide independent estimates of some quantity, say  a project task duration. Further, let us assume that each individual has a probability p of making a correct estimate.

Based on the above, the probability that they both make a correct estimate, P(\textnormal{both correct}),  is:

P(\textnormal{both correct}) = p*p = p^2,

This is a consequence of our assumption that the individual estimates are independent of each other.

Similarly,  the probability that they both get it wrong, P(\textnormal{both wrong}), is:

P(\textnormal{both wrong}) = (1-p)*(1-p) = (1-p)^2,

Now we can ask the following question:

What is the probability that both individuals make the correct estimate if we know that they have both made the same estimate?

This can be figured out using Bayes’ Theorem, which in the context of the question can be stated as follows:

P(\textnormal{both correct\textbar same estimate})= \displaystyle{\frac{ P(\textnormal{same estimate\textbar both correct})*P(\textnormal{both correct})}{ P(\textnormal{same estimate})}}

In the above equation, P(\textnormal{both correct\textbar same estimate}) is the probability that both individuals get it right given that they have made the same estimate (which is  what we want to figure out). This is an example of a conditional probability – i.e.  the probability that an event occurs given that another, possibly related event has already occurred.  See this post for a detailed discussion of conditional probabilities.

Similarly, P(\textnormal{same estimate\textbar both correct}) is the conditional probability that both estimators make the same estimate given that they are both correct. This probability is 1.

Question: Why? 

Answer: If both estimators are correct then they must have made the same estimate (i.e. they must both within be an acceptable range of the right answer).

Finally, P(\textnormal{same estimate}) is the probability that both make the same estimate. This is simply the sum of the probabilities that both get it right and both get it wrong. Expressed in terms of p this is, p^2+(1-p)^2.

Now lets apply Bayes’ theorem to the following two cases:

  1. Both individuals are good estimators – i.e. they have a high probability of making a correct estimate. We’ll assume they both have a 90% chance of getting it right (p=0.9).
  2. Both individuals are poor estimators – i.e. they have a low probability of making a correct estimate. We’ll assume they both have a 30% chance of getting it right (p=0.3)

Consider the first case. The probability that both estimators get it right given that they make the same estimate is:

P(\textnormal{both correct\textbar same estimate})= \displaystyle\frac{1*0.9*0.9}{0.9*0.9+0.1*0.1}= \displaystyle \frac{0.81}{0.82}= 0.9878

Thus we see that the group estimate has a significantly better chance of being right than the individual ones:  a probability of 0.9878 as opposed to 0.9.

In the second case, the probability that both get it right is:

P(\textnormal{both correct\textbar same estimate})= \displaystyle \frac{1*0.3*0.3}{0.3*0.3+0.7*0.7}= \displaystyle \frac{0.09}{0.58}= 0.155

The situation is completely reversed: the group estimate has a much smaller chance of being right than an  individual estimate!

In summary:  estimates provided by a group consisting of individuals of similar ability working independently are more likely to be right (compared to individual estimates) if the group consists of  competent estimators and more likely to be wrong (compared to individual estimates) if the group consists of  poor estimators.

Assumptions and complications

I have made a number of simplifying assumptions in the above argument. I discuss these below with some commentary.

  1. The main assumption is that individuals work independently. This assumption is not valid for many situations. For example, project estimates are often made  by a group of people working together.  Although one can’t work out what will happen in such situations using the arguments of the previous section, it is reasonable to assume that given the right conditions, estimators will use their collective knowledge to work collaboratively.   Other things being equal,  such collaboration would lead a group of skilled estimators to reinforce each others’ estimates (which are likely to be quite similar) whereas less skilled ones may spend time arguing over their (possibly different and incorrect) guesses.  Based on this, it seems reasonable to conjecture that groups consisting of good estimators will tend to make even better estimates than they would individually whereas those consisting of poor estimators have a significant chance of making worse ones.
  2. Another assumption is that an estimate is either good or bad. In reality there is a range that is neither good nor bad, but may be acceptable.
  3. Yet another assumption is that an estimator’s ability can be accurately quantified using a single numerical probability.  This is fine providing the number actually represents the person’s estimation ability for the situation at hand. However, typically such probabilities are evaluated on the basis of  past estimates. The problem is, every situation is unique and history may not be a good guide to the situation at hand. The best way to address this is to involve people with diverse experience in the estimation exercise.  This will almost often lead to  a significant spread of estimates which may then have to be refined by debate and negotiation.

Real-life estimation situations have a number of other complications.  To begin with, the influence that specific individuals have on the estimation process may vary – a manager who is  a poor estimator may, by virtue of his position, have a greater influence than others in a group. This will skew the group estimate by a factor that cannot be estimated.  Moreover, strategic behaviour may influence estimates in a myriad other ways. Then there is the groupthink factor  as well.

…and I’m sure there are many others.

Finally I should mention that group estimates can depend on the details of the estimation process. For example, research suggests that under certain conditions competition can lead to better estimates than cooperation.


In this post I have attempted to make some general inferences regarding the validity of group estimates based on arguments involving conditional probabilities. The arguments suggest that, all other things being equal, a collective estimate from a bunch of skilled estimators will generally be better than their individual estimates whereas an estimate from a group of less skilled estimators will tend to be worse than their individual estimates. Of course, in real life, there are a host of other factors  that can come into play:  power, politics and biases being just a few. Though these are often hidden, they can  influence group estimates in inestimable ways.


Thanks go out to George Gkotsis and Craig Brown for their comments which inspired this post.

Written by K

December 1, 2011 at 5:16 am

The drunkard’s dartboard revisited: yet another Excel-based example of Monte Carlo simulation

with 6 comments

(Note: An Excel sheet showing sample calculations and plots discussed in this post can be downloaded here.)


Some months ago, I wrote a post explaining the basics of Monte Carlo simulation using the example of a drunkard throwing darts at a board. In that post I assumed that the darts could land anywhere on the dartboard with equal probability. In other words, the hit locations were assumed to be uniformly distributed. In a comment on the piece, George Gkotsis challenged this assumption, arguing that that regardless of the level of inebriation of the thrower, a dart would be more likely to land near the centre of the board than away from it (providing the player is at least moderately skilled). He also suggested using the Normal Distribution to model the spread of hits, with the variance of the distribution serving as a rough measure of the inaccuracy (or drunkenness!) of the drunkard. In George’s words:

I would propose to introduce a ‘skill’ factor, which represents the circle/square ratio (maybe a normal-Gaussian distribution). Of course, this skill factor would be very low (high variance) for a drunken player, but would still take into account the fact that throwing darts into a square is not purely random.

In this post I revisit the drunkard’s dartboard, taking into account George’s suggestions.

Setting the stage

To keep things simple, I’ll make the following assumptions:

Figure 1: The dartboard

  1. The dartboard is a circle of radius 0.5 units centred at the origin (see Figure 1)
  2. The chance of a hit is greatest at the centre of the dartboard and falls off as one moves away from it.
  3. The distribution of hits is a function of distance from the centre but does not depend on direction. In mathematical terms, for a given distance r from the centre of the dartboard, the dart can land at any angle \theta with equal probability, \theta being the angle between the line joining the centre of the board to the dart and the x axis. See Figure 2 for graphical representations of a hit location in terms of r and \theta. Note that that the x and y coordinates can be obtained using the formulas x = r\cos\theta and y= r\sin\theta as s shown in Figure 2.
  4. Hits are distributed according to the Normal distribution with maximum at the centre of the dartboard.
  5. The variance of the Normal distribution is a measure of inaccuracy/drunkenness of the drunkard: the more drunk the drunk, the greater the variation in his aim.

Figure 2: The coordinates of a hit location

These assumptions are consistent with George’s suggestions.

The simulation

[Note to the reader: you may want to download the demo before continuing.]

The steps of a simulation run are as follows:

  1. Generate a number that is normally distributed with a zero mean and a specified standard deviation. This gives the distance, r, of a randomly thrown dart from the centre of the board for a player with a “inaccuracy factor” represented by the standard deviation. Column A in the demo contains normally distributed random numbers with zero mean and a standard deviation of 0.2 . Note that I selected the latter number for no other reason than the results show up clearly on a fixed-axis plot shown in Figure 2.
  2. Generate a uniformly distributed random number lying between 0 and 2\pi. This represents the angle \theta. This is the content of column B of the demo.
  3. The numbers obtained from steps 1 and 2 for completely specify the location of a hit. The location’s x and y coordinates can be worked out using the formulas x = r\cos\theta and y= r\sin\theta. These are listed in columns C and D in the Excel demo.
  4. Re-run steps 1 through 4 as many times as needed. Note that the demo is set up for 5000 runs. You can change this manually or, better yet, automate it. The latter is left as an exercise for you.

It is instructive to visualize the resulting hits using a scatter plot. Among other things this can tell you, at a glance, if the results make sense. For example, we would expect hits to be symmetrically distributed about the origin because the drunkard’s throws are not biased in any particular direction around the centre). A non-symmetrical distribution is thus an indication that there is an error in the calculations.

Now, any finite collection of hits is unlikely to be perfectly symmetrical because of outliers. Nevertheless, the distributions should be symmetrical on average. To test this, run the demo a few times (hit F9 with the demo open). Notice how the position of outliers and the overall shape of the distribution of points changes randomly from simulation to simulation. In all cases, however, there is a clear maximum at the centre of the dartboard with the probability of a hit falling with distance from the centre.

Figure 3: Scatter plot for standard deviation=0.2

Figure 3 shows the results of simulations for a standard deviation of 0.2. Figures 4 and 5 show the results of simulations for standard deviations of 0.1 and 0.4.

Figure 4: Scatter plot for standard deviation=0.1

Note that the plot has fixed axes- i.e. the area depicted is the 1×1 square that encloses the dartboard, regardless of the standard deviation. Consequently, for larger standard deviations (such as 0.4) many hits will be out of range and will not show up on the plot.

Figure 5: Scatter plot for standard deviation=0.4

Closing remarks

As I have stressed in my previous posts on Monte Carlo simulation, the usefulness of a simulation depends on the choice of an appropriate distribution. If the selected distribution does not reflect reality, neither will the simulation. This is true regardless of whether one is simulating a drunkard’s wayward aim or the duration of project task. You may have noted that the assumption of normally-distributed hits has no justification whatsoever; it is just as arbitrary as my original assumption of uniformity. In fact, the hit locations of drunken dart throws is highly unlikely to be either uniform or Normal. Nevertheless, I hope that some of my readers will find the above example to be of pedagogical value.


Thanks to George Gkotsis for his comment which got me thinking about this post.

Written by K

November 3, 2011 at 4:59 am

Uncertainty about uncertainty

leave a comment »


More often than not, managerial decisions are made on the basis of uncertain information. To lend some rigour to the process of decision making, it is sometimes assumed that uncertainties of interest can be quantified accurately using probabilities. As it turns out, this assumption can be incorrect in many situations because the probabilities themselves can be uncertain.   In this post I discuss a couple of ways in which such uncertainty about uncertainty can manifest itself.

The problem of vagueness

In a paper entitled, “Is Probability the Only Coherent Approach to Uncertainty?”,  Mark Colyvan made a distinction between two types of uncertainty:

  1. Uncertainty about some underlying fact. For example, we might be uncertain about the cost of a project – that there will be a cost is a fact, but we are uncertain about what exactly it will be.
  2. Uncertainty about situations where there is no underlying fact.  For example, we might be uncertain about whether customers will be satisfied with the outcome of a project. The problem here is the definition of customer satisfaction. How do we measure it? What about customers who are neither satisfied nor dissatisfied?  There is no clear-cut definition of what customer satisfaction actually is.

The first type of uncertainty refers to the lack of knowledge about something that we know exists. This is sometimes referred to as epistemic uncertainty – i.e. uncertainty pertaining to knowledge. Such uncertainty arises from imprecise measurements, changes in the object of interest etc.  The key point is that we know for certain that the item of  interest has well-defined properties, but we don’t know what they are and hence the uncertainty. Such uncertainty can be quantified accurately using probability.

Vagueness, on the other hand, arises from an imprecise use of language.  Specifically, the term refers to the use of criteria that cannot distinguish between borderline cases.  Let’s clarify this using the example discussed earlier.  A popular way to measure customer satisfaction is through surveys. Such surveys may be able to tell us that customer A is more satisfied than customer B. However, they cannot distinguish between borderline cases because any boundary between satisfied and not satisfied customers is arbitrary.  This problem becomes apparent when considering an indifferent customer. How should such a customer be classified – satisfied or not satisfied? Further, what about customers who choose not to respond? It is therefore clear that any numerical probability computed from such data cannot be considered accurate.  In other words, the probability itself is uncertain.

Ambiguity in classification

Although the distinction made by Colyvan is important, there is a deeper issue that can afflict uncertainties that appear to be quantifiable at first sight. To understand how this happens, we’ll first need to take a brief look at how probabilities are usually computed.

An operational definition of probability is that it is the ratio of the number of times the event of interest occurs to the total number of events observed. For example, if my manager notes my arrival times at work over 100 days and finds that I arrive before 8:00 am on 62 days then he could infer that the probability my arriving before 8:00 am is 0.62.   Since the probability is assumed to equal the frequency of occurrence of the event of interest, this is sometimes called the frequentist interpretation of probability.

The above seems straightforward enough, so you might be asking: where’s the problem?

The problem is that events can generally be classified in several different ways and the computed probability of an event occurring can depend on the way that it is classified. This is called the reference class problem.   In a paper entitled, “The Reference Class Problem is Your Problem Too”, Alan Hajek described the reference class problem as follows:

“The reference class problem arises when we want to assign a probability to a proposition (or sentence, or event) X, which may be classified in various ways, yet its probability can change depending on how it is classified.”

Consider the situation I mentioned earlier. My manager’s approach seems reasonable, but there is a problem with it: all days are not the same as far as my arrival times are concerned. For example, it is quite possible that my arrival time is affected by the weather: I may arrive later on rainy days than on sunny ones.  So, to get a better estimate my manager should also factor in the weather. He would then end up with two probabilities, one for fine weather and the other for foul. However, that is not all: there are a number of other criteria that could affect my arrival times – for example, my state of health (I may call in sick and not come in to work at all), whether I worked late the previous day etc.

What seemed like a straightforward problem is no longer so because of the uncertainty regarding which reference class is the right one to use.

Before closing this section, I should mention that the reference class problem has implications for many professional disciplines. I have discussed its relevance to project management in my post entitled, “The reference class problem and its implications for project management”.

To conclude

In this post we have looked at a couple of forms of uncertainty about uncertainty that have practical implications for decision makers. In particular, we have seen that probabilities used in managerial decision making can be uncertain because of  vague definitions of events and/or ambiguities in their classification.  The bottom line for those who use probabilities to support decision-making is to ensure that the criteria used to determine events of interest refer to unambiguous facts that are appropriate to the situation at hand.  To sum up: decisions made on the basis of probabilities are only as good as the assumptions that go into them, and the assumptions themselves may be prone to uncertainties such as the ones described in this article.

Written by K

September 29, 2011 at 10:34 pm

The drunkard’s dartboard: an intuitive explanation of Monte Carlo methods

with 10 comments

(Note to the reader: An Excel sheet showing sample calculations and plots discussed in this post  can be downloaded here.)

Monte Carlo simulation techniques have been applied to areas ranging from physics to project management. In earlier posts, I discussed how these methods can be used to simulate project task durations (see this post and this one for example). In those articles, I described simulation procedures in enough detail for readers to be able to reproduce the calculations for themselves. However, as my friend Paul Culmsee mentioned, the mathematics tends to obscure the rationale behind the technique. Indeed, at first sight it seems somewhat paradoxical that one can get accurate answers via random numbers. In this post, I illustrate the basic idea behind Monte Carlo methods through an example that involves nothing more complicated than squares and circles. To begin with, however, I’ll start with something even simpler – a drunken darts player.

Consider a sozzled soul who is throwing darts at a board situated some distance from him.  To keep things simple, we’ll assume the following:

  1. The board is modeled by the circle shown in Figure 1, and our souse scores a point if the dart falls within the circle.
  2. The dart board is inscribed in a square with sides 1 unit long as shown in the figure, and we’ll assume for simplicity that the  dart always  falls somewhere  within the square (our protagonist is not that smashed).
  3. Given his state, our hero’s aim is not quite what it should be –  his darts fall anywhere within the square with equal probability. (Note added on 01 March 2011: See the comment by George Gkotsis below for a critique of this assumption)

Figure 1: "Dartboard" and enclosing square

We can simulate the results of our protagonist’s unsteady attempts by generating two sets of  uniformly distributed random numbers lying between 0 and 1 (This is easily done in Excel using the rand() function).  The pairs of random numbers thus generated – one from each set –  can be treated as the (x,y) coordinates of  the dart for a particular throw. The result of 1000 pairs of random numbers thus generated (representing the drunkard’s dart throwing attempts) is shown in Figure 2 (For those interested in seeing the details, an Excel sheet showing the calculations for 100o trials can be downloaded here).

Figure 2: Result for 1000 throws

A trial results in a “hit” if it lies within the circle.  That is, if it satisfies the following equation:

(x-0.5)^2 + (y-0.5)^2 < 0.25\ldots \ldots (1)

(Note:  if we replace “<”  by “=”  in the above expression, we get the equation for a circle of radius 0.5 units, centered at x=0.5 and y=0.5.)

Now, according to the frequency interpretation of probability, the probability of the plastered player scoring a point is approximated by the ratio of the number of hits in the circle to the total number of attempts. In this case, I get an average of 790/1000 which is 0.79 (generated from 10 sets of 1000 trials each). Your result will be different from because you will generate different sets of random numbers from the ones I did. However, it should be reasonably close to my result.

Further, the frequency interpretation of probability tells us that the approximation becomes more accurate as the number of trials increases. To see why this is so, let’s increase the number of trials and plot the results. I carried out simulations for 2000, 4000, 8000 and 16000 trials. The results of these simulations are shown in Figures 3 through 6.

Figure 3: Result for 2000 throws

Figure 4: Result for 4000 throws

Figure 5: Result for 8000 throws

Figure 6: Result for 16000 throws

Since a dart is equally likely to end up anywhere within the square, the exact probability of a hit is simply the area of the dartboard (i.e. the circle)  divided by the entire area over which the dart can land. In this case, since the area of the enclosure (where the dart must fall) is 1 square unit, the area of the dartboard is actually equal to the probability.  This is easily seen by calculating the area of the circle using the standard formula \pi r^2 where r is the radius of the circle (0.5 units in this case). This yields 0.785398 sq units, which is reasonably close to the number that we got for the 1000 trial case.  In the 16000 trial case, I  get a number that’s closer to the exact result: an average of 0.7860 from 10 sets of 16000 trials.

As we see from Figure 6, in the 16000 trial case, the entire square is peppered with closely-spaced “dart marks” – so much so, that it looks as though the square is a uniform blue.  Hence, it seems intuitively clear that as we increase, the number of throws, we should get a better approximation of the area and, hence, the probability.

There are a couple of points worth mentioning here. First, in principle this technique can be used to calculate areas of figures of any shape. However, the more irregular the figure, the worse the approximation – simply because it becomes less likely that the entire figure will be sampled correctly by “dart throws.” Second,  the reader may have noted that although the 16000 trial case gives a good enough result for the area of the circle, it isn’t particularly accurate considering the large number of trials. Indeed, it is known that the “dart approximation” is not a very good way of calculating areas – see this note for more on this point.

Finally, let’s look at connection between the general approach used in Monte Carlo techniques and the example discussed above  (I use the steps described in the Wikipedia article on Monte Carlo methods as representative of the general approach):

  1. Define a domain of possible inputs – in our case the domain of inputs is defined by the enclosing square of side 1 unit.
  2. Generate inputs randomly from the domain using a certain specified probability distribution – in our example the probability distribution is a pair of independent, uniformly distributed random numbers lying between 0 and 1.
  3. Perform a computation using the inputs – this is the calculation that determines whether or not a particular trial is a hit or not (i.e.  if the x,y coordinates obey inequality  (1) it is a hit, else  it’s a miss)
  4. Aggregate the results of the individual computations into the final result – This corresponds to the calculation of the probability (or equivalently, the area of the circle) by aggregating the number of hits for each set of trials.

To summarise: Monte Carlo algorithms generate random variables (such as probability) according to pre-specified distributions.  In most practical applications  one will use more efficient techniques to sample the distribution (rather than the naïve method I’ve used here.)  However, the basic idea is as simple as playing drunkard’s darts.


Thanks go out to Vlado Bokan for helpful conversations while this post was being written and to Paul Culmsee for getting me thinking about a simple way to explain Monte Carlo methods.

Written by K

February 25, 2011 at 4:51 am

%d bloggers like this: