Archive for the ‘Data Analytics’ Category
Most machine learning algorithms involve minimising an error measure of some kind (this measure is often called an objective function or loss function). For example, the error measure in linear regression problems is the famous mean squared error – i.e. the averaged sum of the squared differences between the predicted and actual values. Like the mean squared error, most objective functions depend on all points in the training dataset. In this post, I describe the support vector machine (SVM) approach which focuses instead on finding the optimal separation boundary between datapoints that have different classifications. I’ll elaborate on what this means in the next section.
Here’s the plan in brief. I’ll begin with the rationale behind SVMs using a simple case of a binary (two class) dataset with a simple separation boundary (I’ll clarify what “simple” means in a minute). Following that, I’ll describe how this can be generalised to datasets with more complex boundaries. Finally, I’ll work through a couple of examples in R, illustrating the principles behind SVMs. In line with the general philosophy of my “Gentle Introduction to Data Science Using R” series, the focus is on developing an intuitive understanding of the algorithm along with a practical demonstration of its use through a toy example.
The basic idea behind SVMs is best illustrated by considering a simple case: a set of data points that belong to one of two classes, red and blue, as illustrated in figure 1 below. To make things simpler still, I have assumed that the boundary separating the two classes is a straight line, represented by the solid green line in the diagram. In the technical literature, such datasets are called linearly separable.
In the linearly separable case, there is usually a fair amount of freedom in the way a separating line can be drawn. Figure 2 illustrates this point: the two broken green lines are also valid separation boundaries. Indeed, because there is a non-zero distance between the two closest points between categories, there are an infinite number of possible separation lines. This, quite naturally, raises the question as to whether it is possible to choose a separation boundary that is optimal.
The short answer is, yes there is. One way to do this is to select a boundary line that maximises the margin, i.e. the distance between the separation boundary and the points that are closest to it. Such an optimal boundary is illustrated by the black brace in Figure 3. The really cool thing about this criterion is that the location of the separation boundary depends only on the points that are closest to it. This means, unlike other classification methods, the classifier does not depend on any other points in dataset. The directed lines between the boundary and the closest points on either side are called support vectors (these are the solid black lines in figure 3). A direct implication of this is that the fewer the support vectors, the better the generalizability of the boundary.
Although the above sounds great, it is of limited practical value because real data sets are seldom (if ever) linearly separable.
So, what can we do when dealing with real (i.e. non linearly separable) data sets?
A simple approach to tackle small deviations from linear separability is to allow a small number of points (those that are close to the boundary) to be misclassified. The number of possible misclassifications is governed by a free parameter C, which is called the cost. The cost is essentially the penalty associated with making an error: the higher the value of C, the less likely it is that the algorithm will misclassify a point.
This approach – which is called soft margin classification – is illustrated in Figure 4. Note the points on the wrong side of the separation boundary. We will demonstrate soft margin SVMs in the next section. (Note: At the risk of belabouring the obvious, the purely linearly separable case discussed in the previous para is simply is a special case of the soft margin classifier.)
Real life situations are much more complex and cannot be dealt with using soft margin classifiers. For example, as shown in Figure 5, one could have widely separated clusters of points that belong to the same classes. Such situations, which require the use of multiple (and nonlinear) boundaries, can sometimes be dealt with using a clever approach called the kernel trick.
The kernel trick
Recall that in the linearly separable (or soft margin) case, the SVM algorithm works by finding a separation boundary that maximises the margin, which is the distance between the boundary and the points closest to it. The distance here is the usual straight line distance between the boundary and the closest point(s). This is called the Euclidean distance in honour of the great geometer of antiquity. The point to note is that this process results in a separation boundary that is a straight line, which as Figure 5 illustrates, does not always work. In fact in most cases it won’t.
So what can we do? To answer this question, we have to take a bit of a detour…
What if we were able to generalize the notion of distance in a way that generates nonlinear separation boundaries? It turns out that this is possible. To see how, one has to first understand how the notion of distance can be generalized.
The key properties that any measure of distance must satisfy are:
- Non-negativity – a distance cannot be negative, a point that needs no further explanation I reckon 🙂
- Symmetry – that is, the distance between point A and point B is the same as the distance between point B and point A.
- Identity– the distance between a point and itself is zero.
- Triangle inequality – that is the sum of distances between point A and B and points B and C must be less than or equal to the distance between A and C (equality holds only if all three points lie along the same line).
Any mathematical object that displays the above properties is akin to a distance. Such generalized distances are called metrics and the mathematical space in which they live is called a metric space. Metrics are defined using special mathematical functions designed to satisfy the above conditions. These functions are known as kernels.
The essence of the kernel trick lies in mapping the classification problem to a metric space in which the problem is rendered separable via a separation boundary that is simple in the new space, but complex – as it has to be – in the original one. Generally, the transformed space has a higher dimensionality, with each of the dimensions being (possibly complex) combinations of the original problem variables. However, this is not necessarily a problem because in practice one doesn’t actually mess around with transformations, one just tries different kernels (the transformation being implicit in the kernel) and sees which one does the job. The check is simple: we simply test the predictions resulting from using different kernels against a held out subset of the data (as one would for any machine learning algorithm).
It turns out that a particular function – called the radial basis function kernel (RBF kernel) – is very effective in many cases. The RBF kernel is essentially a Gaussian (or Normal) function with the Euclidean distance between pairs of points as the variable (see equation 1 below). The basic rationale behind the RBF kernel is that it creates separation boundaries that it tends to classify points close together (in the Euclidean sense) in the original space in the same way. This is reflected in the fact that the kernel decays (i.e. drops off to zero) as the Euclidean distance between points increases.
The rate at which a kernel decays is governed by the parameter – the higher the value of , the more rapid the decay. This serves to illustrate that the RBF kernel is extremely flexible….but the flexibility comes at a price – the danger of overfitting for large values of . One should choose appropriate values of C and so as to ensure that the resulting kernel represents the best possible balance between flexibility and accuracy. We’ll discuss how this is done in practice later in this article.
Finally, though it is probably obvious, it is worth mentioning that the separation boundaries for arbitrary kernels are also defined through support vectors as in Figure 3. To reiterate a point made earlier, this means that a solution that has fewer support vectors is likely to be more robust than one with many. Why? Because the data points defining support vectors are ones that are most sensitive to noise- therefore the fewer, the better.
There are many other types of kernels, each with their own pros and cons. However, I’ll leave these for adventurous readers to explore by themselves. Finally, for a much more detailed….and dare I say, better… explanation of the kernel trick, I highly recommend this article by Eric Kim.
Support vector machines in R
In this demo we’ll use the svm interface that is implemented in the e1071 R package. This interface provides R programmers access to the comprehensive libsvm library written by Chang and Lin. I’ll use two toy datasets: the famous iris dataset available with the base R package and the sonar dataset from the mlbench package. I won’t describe details of the datasets as they are discussed at length in the documentation that I have linked to. However, it is worth mentioning the reasons why I chose these datasets:
- As mentioned earlier, no real life dataset is linearly separable, but the iris dataset is almost so. Consequently, it is a good illustration of using linear SVMs. Although one almost never uses these in practice, I have illustrated their use primarily for pedagogical reasons.
- The sonar dataset is a good illustration of the benefits of using RBF kernels in cases where the dataset is hard to visualise (60 variables in this case!). In general, one would almost always use RBF (or other nonlinear) kernels in practice.
With that said, let’s get right to it. I assume you have R and RStudio installed. For instructions on how to do this, have a look at the first article in this series. The processing preliminaries – loading libraries, data and creating training and test datasets are much the same as in my previous articles so I won’t dwell on these here. For completeness, however, I’ll list all the code so you can run it directly in R or R studio (a complete listing of the code can be found here):
The output from the SVM model show that there are 24 support vectors. If desired, these can be examined using the SV variable in the model – i.e via svm_model$SV.
The test prediction accuracy indicates that the linear performs quite well on this dataset, confirming that it is indeed near linearly separable. To check performance by class, one can create a confusion matrix as described in my post on random forests. I’ll leave this as an exercise for you. Another point is that we have used a soft-margin classification scheme with a cost C=1. You can experiment with this by explicitly changing the value of C. Again, I’ll leave this for you an exercise.
Before proceeding to the RBF kernel, I should mention a point that an alert reader may have noticed. The predicted variable, Species, can take on 3 values (setosa, versicolor and virginica). However, our discussion above dealt with a binary (2 valued) classification problem. This brings up the question as to how the algorithm deals multiclass classification problems – i.e those involving datasets with more than two classes. The libsvm algorithm (which svm uses) does this using a one-against-one classification strategy. Here’s how it works:
- Divide the dataset (assumed to have N classes) into N(N-1)/2 datasets that have two classes each.
- Solve the binary classification problem for each of these subsets
- Use a simple voting mechanism to assign a class to each data point.
Basically, each data point is assigned the most frequent classification it receives from all the binary classification problems it figures in.
With that said for the unrealistic linear classifier, let’s move to the real world. In the code below, I build SVM models using three different kernels
- Linear kernel (this is for comparison with the following 2 kernels).
- RBF kernel with default values for the parameters and .
- RBF kernel with optimal values for and . The optimal values are obtained using the tune.svm function (also available in e1071), which essentially builds models for multiple combinations of parameter values and selects the best.
OK, lets go:
I’ll leave you to examine the contents of the model. The important point to note here is that the performance of the model with the test set is quite dismal compared to the previous case. This simply indicates that the linear kernel is not appropriate here. Let’s take a look at what happens if we use the RBF kernel with default values for the parameters:
That’s a pretty decent improvement from the linear kernel. Let’s see if we can do better by doing some parameter tuning. To do this we first invoke tune.svm and use the parameters it gives us in the call to svm:
Which is fairly decent improvement on the un-optimised case.
This bring us to the end of this introductory exploration of SVMs in R. To recap, the distinguishing feature of SVMs in contrast to most other techniques is that they attempt to construct optimal separation boundaries between different categories.
SVMs are quite versatile and have been applied to a wide variety of domains ranging from chemistry to pattern recognition. They are best used in binary classification scenarios. This brings up a question as to where SVMs are to be preferred to other binary classification techniques such as logistic regression. The honest response is, “it depends” – but here are some points to keep in mind when choosing between the two. A general point to keep in mind is that SVM algorithms tend to be expensive both in terms of memory and computation, issues that can start to hurt as the size of the dataset increases.
Given all the above caveats and considerations, the best way to figure out whether an SVM approach will work for your problem may be to do what most machine learning practitioners do: try it out!
Data scientists are sometimes blind to the possibility that the predictions of their algorithms can have unforeseen negative effects on people. Ethical or social implications are easy to overlook when one finds interesting new patterns in data, especially if they promise significant financial gains. The Centrelink debt recovery debacle, recently reported in the Australian media, is a case in point.
Here is the story in brief:
Centrelink is an Australian Government organisation responsible for administering welfare services and payments to those in need. A major challenge such organisations face is ensuring that their clients are paid no less and no more than what is due to them. This is difficult because it involves crosschecking client income details across multiple systems owned by different government departments, a process that necessarily involves many assumptions. In July 2016, Centrelink unveiled an automated compliance system that compares income self-reported by clients to information held by the taxation office.
The problem is that the algorithm is flawed: it makes strong (and incorrect!) assumptions regarding the distribution of income across a financial year and, as a consequence, unfairly penalizes a number of legitimate benefit recipients. It is very likely that the designers and implementers of the algorithm did not fully understand the implications of their assumptions. Worse, from the errors made by the system, it appears they may not have adequately tested it either. But this did not stop them (or, quite possibly, their managers) from unleashing their algorithm on an unsuspecting public, causing widespread stress and distress. More on this a bit later.
Algorithms like the one described above are the subject of Cathy O’Neil’s aptly titled book, Weapons of Math Destruction. In the remainder of this article I discuss the main themes of the book. Just to be clear, this post is more riff than review. However, for those seeking an opinion, here’s my one-line version: I think the book should be read not only by data science practitioners, but also by those who use or are affected by their algorithms (which means pretty much everyone!).
Abstractions and assumptions
‘O Neil begins with the observation that data algorithms are mathematical models of reality, and are necessarily incomplete because several simplifying assumptions are invariably baked into them. This point is important and often overlooked so it is worth illustrating via an example.
When assessing a person’s suitability for a loan, a bank will want to know whether the person is a good risk. It is impossible to model creditworthiness completely because we do not know all the relevant variables and those that are known may be hard to measure. To make up for their ignorance, data scientists typically use proxy variables, i.e. variables that are believed to be correlated with the variable of interest and are also easily measurable. In the case of creditworthiness, proxy variables might be things like gender, age, employment status, residential postcode etc. Unfortunately many of these can be misleading, discriminatory or worse, both.
The Centrelink algorithm provides a good example of such a “double-whammy” proxy. The key variable it uses is the difference between the client’s annual income reported by the taxation office and self-reported annual income stated by the client. A large difference is taken to be an indicative of an incorrect payment and hence an outstanding debt. This simplistic assumption overlooks the fact that most affected people are not in steady jobs and therefore do not earn regular incomes over the course of a financial year (see this article by Michael Griffin, for a detailed example). Worse, this crude proxy places an unfair burden on vulnerable individuals for whom casual and part time work is a fact of life.
Worse still, for those wrongly targeted with a recovery notice, getting the errors sorted out is not a straightforward process. This is typical of a WMD. As ‘O Neil states in her book, “The human victims of WMDs…are held to a far higher standard of evidence than the algorithms themselves.” Perhaps this is because the algorithms are often opaque. But that’s a poor excuse. This is the only technical field where practitioners are held to a lower standard of accountability than those affected by their products.
‘O Neil’s sums it up rather nicely when she calls algorithms like the Centrelink one weapons of math destruction (WMD).
Self-fulfilling prophecies and feedback loops
A characteristic of WMD is that their predictions often become self-fulfilling prophecies. For example a person denied a loan by a faulty risk model is more likely to be denied again when he or she applies elsewhere, simply because it is on their record that they have been refused credit before. This kind of destructive feedback loop is typical of a WMD.
An example that ‘O Neil dwells on at length is a popular predictive policing program. Designed for efficiency rather than nuanced judgment, such algorithms measure what can easily be measured and act by it, ignoring the subtle contextual factors that inform the actions of experienced officers on the beat. Worse, they can lead to actions that can exacerbate the problem. For example, targeting young people of a certain demographic for stop and frisk actions can alienate them to a point where they might well turn to crime out of anger and exasperation.
As Goldratt famously said, “Tell me how you measure me and I’ll tell you how I’ll behave.”
This is not news: savvy managers have known about the dangers of managing by metrics for years. The problem is now exacerbated manyfold by our ability to implement and act on such metrics on an industrial scale, a trend that leads to a dangerous devaluation of human judgement in areas where it is most needed.
A related problem – briefly mentioned earlier – is that some of the important variables are known but hard to quantify in algorithmic terms. For example, it is known that community-oriented policing, where officers on the beat develop relationships with people in the community, leads to greater trust. The degree of trust is hard to quantify, but it is known that communities that have strong relationships with their police departments tend to have lower crime rates than similar communities that do not. Such important but hard-to-quantify factors are typically missed by predictive policing programs.
Ironically, although WMDs can cause destructive feedback loops, they are often not subjected to feedback themselves. O’Neil gives the example of algorithms that gauge the suitability of potential hires. These programs often use proxy variables such as IQ test results, personality tests etc. to predict employability. Candidates who are rejected often do not realise that they have been screened out by an algorithm. Further, it often happens that candidates who are thus rejected go on to successful careers elsewhere. However, this post-rejection information is never fed back to the algorithm because it impossible to do so.
In such cases, the only way to avoid being blackballed is to understand the rules set by the algorithm and play according to them. As ‘O Neil so poignantly puts it, “our lives increasingly depend on our ability to make our case to machines.” However, this can be difficult because it assumes that a) people know they are being assessed by an algorithm and 2) they have knowledge of how the algorithm works. In most hiring scenarios neither of these hold.
Just to be clear, not all data science models ignore feedback. For example, sabermetric algorithms used to assess player performance in Major League Baseball are continually revised based on latest player stats, thereby taking into account changes in performance.
Driven by data
In recent years, many workplaces have gradually seen the introduction to data-driven efficiency initiatives. Automated rostering, based on scheduling algorithms is an example. These algorithms are based on operations research techniques that were developed for scheduling complex manufacturing processes. Although appropriate for driving efficiency in manufacturing, these techniques are inappropriate for optimising shift work because of the effect they have on people. As O’ Neil states:
Scheduling software can be seen as an extension of just-in-time economy. But instead of lawn mower blades or cell phone screens showing up right on cue, it’s people, usually people who badly need money. And because they need money so desperately, the companies can bend their lives to the dictates of a mathematical model.
She correctly observes that an, “oversupply of low wage labour is the problem.” Employers know they can get away with treating people like machine parts because they have a large captive workforce. What makes this seriously scary is that vested interests can make it difficult to outlaw such exploitative practices. As ‘O Neil mentions:
Following [a] New York Times report on Starbucks’ scheduling practices, Democrats in Congress promptly drew up bills to rein in scheduling software. But facing a Republican majority fiercely opposed to government regulations, the chances that their bill would become law were nil. The legislation died.
Commercial interests invariably trump social and ethical issues, so it is highly unlikely that industry or government will take steps to curb the worst excesses of such algorithms without significant pressure from the general public. A first step towards this is to educate ourselves on how these algorithms work and the downstream social effects of their predictions.
Messing with your mind
There is an even more insidious way that algorithms mess with us. Hot on the heels of the recent US presidential election, there were suggestions that fake news items on Facebook may have influenced the results. Mark Zuckerberg denied this, but as this Casey Newton noted in this trenchant tweet, the denial leaves Facebook in “the awkward position of having to explain why they think they drive purchase decisions but not voting decisions.”
Be that as it may, the fact is Facebook’s own researchers have been conducting experiments to fine tune a tool they call the “voter megaphone”. Here’s what ‘O Neil says about it:
The idea was to encourage people to spread the word that they had voted. This seemed reasonable enough. By sprinkling people’s news feeds with “I voted” updates, Facebook was encouraging Americans – more that sixty-one million of them – to carry out their civic duty….by posting about people’s voting behaviour, the site was stoking peer pressure to vote. Studies have shown that the quiet satisfaction of carrying out a civic duty is less likely to move people than the possible judgement of friends and neighbours…The Facebook started out with a constructive and seemingly innocent goal to encourage people to vote. And it succeeded…researchers estimated that their campaign had increased turnout by 340,000 people. That’s a big enough crowd to swing entire states, and even national elections.
And if that’s not scary enough, try this:
For three months leading up to the election between President Obama and Mitt Romney, a researcher at the company….altered the news feed algorithm for about two million people, all of them politically engaged. The people got a higher proportion of hard news, as opposed to the usual cat videos, graduation announcements, or photos from Disney world….[the researcher] wanted to see if getting more [political] news from friends changed people’s political behaviour. Following the election [he] sent out surveys. The self-reported results that voter participation in this group inched up from 64 to 67 percent.
This might not sound like much, but considering the thin margins of recent presidential elections, it could be enough to change a result.
But it’s even more insidious. In a paper published in 2014, Facebook researchers showed that users’ moods can be influenced by the emotional content of their newsfeeds. Here’s a snippet from the abstract of the paper:
In an experiment with people who use Facebook, we test whether emotional contagion occurs outside of in-person interaction between individuals by reducing the amount of emotional content in the News Feed. When positive expressions were reduced, people produced fewer positive posts and more negative posts; when negative expressions were reduced, the opposite pattern occurred. These results indicate that emotions expressed by others on Facebook influence our own emotions, constituting experimental evidence for massive-scale contagion via social networks.
As you might imagine, there was a media uproar following which the lead researcher issued a clarification and Facebook officials duly expressed regret (but, as far as I know, not an apology). To be sure, advertisers have been exploiting this kind of “mind control” for years, but a public social media platform should (expect to) be held to a higher standard of ethics. Facebook has since reviewed its internal research practices, but the recent fake news affair shows that the story is to be continued.
Disarming weapons of math destruction
The Centrelink debt debacle, Facebook mood contagion experiments and the other case studies mentioned in the book illusrate the myriad ways in which Big Data algorithms have a pernicious effect on our day-to-day lives. Quite often people remain unaware of their influence, wondering why a loan was denied or a job application didn’t go their way. Just as often, they are aware of what is happening, but are powerless to change it – shift scheduling algorithms being a case in point.
This is not how it was meant to be. Technology was supposed to make life better for all, not just the few who wield it.
So what can be done? Here are some suggestions:
- To begin with, education is the key. We must work to demystify data science, create a general awareness of data science algorithms and how they work. O’ Neil’s book is an excellent first step in this direction (although it is very thin on details of how the algorithms work)
- Develop a code of ethics for data science practitioners. It is heartening to see that IEEE has recently come up with a discussion paper on ethical considerations for artificial intelligence and autonomous systems and ACM has proposed a set of principles for algorithmic transparency and accountability. However, I should also tag this suggestion with the warning that codes of ethics are not very effective as they can be easily violated. One has to – somehow – embed ethics in the DNA of data scientists. I believe, one way to do this is through practice-oriented education in which data scientists-in-training grapple with ethical issues through data challenges and hackathons. It is as Wittgenstein famously said, “it is clear that ethics cannot be articulated.” Ethics must be practiced.
- Put in place a system of reliable algorithmic audits within data science departments, particularly those that do work with significant social impact.
- Increase transparency a) by publishing information on how algorithms predict what they predict and b) by making it possible for those affected by the algorithm to access the data used to classify them as well as their classification, how it will be used and by whom.
- Encourage the development of algorithms that detect bias in other algorithms and correct it.
- Inspire aspiring data scientists to build models for the good.
It is only right that the last word in this long riff should go to ‘O Neil whose work inspired it. Towards the end of her book she writes:
Big Data processes codify the past. They do not invent the future. Doing that requires moral imagination, and that’s something that only humans can provide. We have to explicitly embed better values into our algorithms, creating Big Data models that follow our ethical lead. Sometimes that will mean putting fairness ahead of profit.
Excellent words for data scientists to live by.
In a previous post, I described how decision tree algorithms work and demonstrated their use via the rpart library in R. Decision trees work by splitting a dataset recursively. That is, subsets arising from a split are further split until a predetermined termination criterion is reached. At each step, a split is made based on the independent variable that results in the largest possible reduction in heterogeneity of the dependent variable.
(Note: readers unfamiliar with decision trees may want to read that post before proceeding)
The main drawback of decision trees is that they are prone to overfitting. The reason for this is that trees, if grown deep, are able to fit all kinds of variations in the data, including noise. Although it is possible to address this partially by pruning, the result often remains less than satisfactory. This is because makes a locally optimal choice at each split without any regard to whether the choice made is the best one overall. A poor split made in the initial stages can thus doom the model, a problem that cannot be fixed by post-hoc pruning.
In this post I describe random forests, a tree-based algorithm that addresses the above shortcoming of decision trees. I’ll first describe the intuition behind the algorithm via an analogy and then do a demo using the R randomForest library.
Motivating random forests
One of the reasons for the popularity of decision trees is that they reflect the way humans make decisions: by weighing up options at each stage and choosing the best one available. The analogy is particularly useful because it also suggests how decision trees can be improved.
One of the lifelines in the game show, Who Wants to be A Millionaire, is “Ask The Audience” wherein a contestant can ask the audience to vote on the answer to a question. The rationale here is that the majority response from a large number of independent decision makers is more likely to yield a correct answer than one from a randomly chosen person. There are two factors at play here:
- People have different experiences and will therefore draw upon different “data” to answer the question.
- People have different knowledge bases and preferences and will therefore draw upon different “variables” to make their choices at each stage in their decision process.
Taking a cue from the above, it seems reasonable to build many decision trees using:
- Different sets of training data.
- Randomly selected subsets of variables at each split of every decision tree.
Predictions can then made by taking the majority vote over all trees (for classification problems) or averaging results over all trees (for regression problems). This is essentially how the random forest algorithm works.
The net effect of the two strategies is to reduce overfitting by a) averaging over trees created from different samples of the dataset and b) decreasing the likelihood of a small set of strong predictors dominating the splits. The price paid is reduced interpretability as well as increased computational complexity. But then, there is no such thing as a free lunch.
The mechanics of the algorithm
Although we will not delve into the mathematical details of the algorithm, it is important to understand how two points made above are implemented in the algorithm.
Bootstrap aggregating… and a (rather cool) error estimate
A key feature of the algorithm is the use of multiple datasets for training individual decision trees. This is done via a neat statistical trick called bootstrap aggregating (also called bagging).
Here’s how bagging works:
Assume you have a dataset of size N. From this you create a sample (i.e. a subset) of size n (n less than or equal to N) by choosing n data points randomly with replacement. “Randomly” means every point in the dataset is equally likely to be chosen and “with replacement” means that a specific data point can appear more than once in the subset. Do this M times to create M equally-sized samples of size n each. It can be shown that this procedure, which statisticians call bootstrapping, is legit when samples are created from large datasets – that is, when N is large.
Because a bagged sample is created by selection with replacement, there will generally be some points that are not selected. In fact, it can be shown that, on the average, each sample will use about two-thirds of the available data points. This gives us a clever way to estimate the error as part of the process of model building.
For every data point, obtain predictions for trees in which the point was out of bag. From the result mentioned above, this will yield approximately M/3 predictions per data point (because a third of the data points are out of bag). Take the majority vote of these M/3 predictions as the predicted value for the data point. One can do this for the entire dataset. From these out of bag predictions for the whole dataset, we can estimate the overall error by computing a classification error (Count of correct predictions divided by N) for classification problems or the root mean squared error for regression problems. This means there is no need to have a separate test data set, which is kind of cool. However, if you have enough data, it is worth holding out some data for use as an independent test set. This is what we’ll do in the demo later.
Using subsets of predictor variables
Although bagging reduces overfitting somewhat, it does not address the issue completely. The reason is that in most datasets a small number of predictors tend to dominate the others. These predictors tend to be selected in early splits and thus influence the shapes and sizes of a significant fraction of trees in the forest. That is, strong predictors enhance correlations between trees which tends to come in the way of variance reduction.
A simple way to get around this problem is to use a random subset of variables at each split. This avoids over-representation of dominant variables and thus creates a more diverse forest. This is precisely what the random forest algorithm does.
Random forests in R
In what follows, I use the famous Glass dataset from the mlbench library. The dataset has 214 data points of six types of glass with varying metal oxide content and refractive indexes. I’ll first build a decision tree model based on the data using the rpart library (recursive partitioning) that I covered in an earlier article and then use then show how one can build a random forest model using the randomForest library. The rationale behind this is to compare the two models – single decision tree vs random forest. In the interests of space, I won’t explain details of the rpart here as I’ve covered it at length in the previous article. However, for completeness, I’ll list the demo code for it before getting into random forests.
Decision trees using rpart
Here’s the code listing for building a decision tree using rpart on the Glass dataset (please see my previous article for a full explanation of each step). Note that I have not used pruning as there is little benefit to be gained from it (Exercise for the reader: try this for yourself!).
Now, we know that decision tree algorithms tend to display high variance so the hit rate from any one tree is likely to be misleading. To address this we’ll generate a bunch of trees using different training sets (via random sampling) and calculate an average hit rate and spread (or standard deviation).
The decision tree algorithm gets it right about 69% of the time with a variation of about 5%. The variation isn’t too bad here, but the accuracy has hardly improved at all (Exercise for the reader: why?). Let’s see if we can do better using random forests.
As discussed earlier, a random forest algorithm works by averaging over multiple trees using bootstrapped samples. Also, it reduces the correlation between trees by splitting on a random subset of predictors at each node in tree construction. The key parameters for randomForest algorithm are the number of trees (ntree) and the number of variables to be considered for splitting (mtry). The algorithm sets a default of 500 for ntree and sets mtry to one-third the total number of predictors for classification problems and square root of the the number of predictors for regression. These defaults can be overridden by explicitly providing values for these variables.
The preliminary stuff – the creation of training and test datasets etc. – is much the same as for decision trees but I’ll list the code for completeness.
randomForest(formula = Type ~ ., data = trainGlass, importance = TRUE, xtest = testGlass[, -typeColNum], ntree = 1001)
The first thing to note is the out of bag error estimate is ~ 24%. Equivalently the hit rate is 76%, which is better than the 69% for decision trees. Secondly, you’ll note that the algorithm does a terrible job identifying type 3 and 6 glasses correctly. This could possibly be improved by a technique called boosting, which works by iteratively improving poor predictions made in earlier stages. I plan to look at boosting in a future post, but if you’re curious, check out the gbm package in R.
Finally, for completeness, let’s see how the test set does:
The test accuracy is better than the out of bag accuracy and there are some differences in the class errors as well. However, overall the two compare quite well and are significantly better than the results of the decision tree algorithm.
Random forest algorithms also give measures of variable importance. Computation of these is enabled by setting importance, a boolean parameter, to TRUE. The algorithm computes two measures of variable importance: mean decrease in Gini and mean decrease in accuracy. Brief explanations of these follow.
Mean decrease in Gini
When determining splits in individual trees, the algorithm looks for the largest class (in terms of population) and attempts to isolate it first. If this is not possible, it tries to do the best it can, always focusing on isolating the largest remaining class in every split.This is called the Gini splitting rule (see this article for a good explanation of the rule).
The “goodness of split” is measured by the Gini Impurity, . For a set containing K categories this is given by:
where is the fraction of the set that belongs to the ith category. Clearly, is 0 when the set is homogeneous or pure (1 class only) and is maximum when classes are equiprobable (for example, in a two class set the maximum occurs when and are 0.5). At each stage the algorithm chooses to split on the predictor that leads to the largest decrease in . The algorithm tracks this decrease for each predictor for all splits and all trees in the forest. The average is reported as the mean decrease in Gini.
Mean decrease in accuracy
The mean decrease in accuracy is calculated using the out of bag data points for each tree. The procedure goes as follows: when a particular tree is grown, the out of bag points are passed down the tree and the prediction accuracy (based on all out of bag points) recorded . The predictors are then randomly permuted and the out of bag prediction accuracy recalculated. The decrease in accuracy for a given predictor is the difference between the accuracy of the original (unpermuted) tree and the those obtained from the permuted trees in which the predictor was excluded. As in the previous case, the decrease in accuracy for each predictor can be computed and tracked as the algorithm progresses. These can then be averaged by predictor to yield a mean decrease in accuracy.
Variable importance plot
From the above, it would seem that the mean decrease in accuracy is a more global measure as it uses fully constructed trees in contrast to the Gini measure which is based on individual splits. In practice, however, there could be other reasons for choosing one over the other…but that is neither here nor there, if you set importance to TRUE, you’ll get both. The numerical measures of importance are returned in the randomForest object (Glass.rf in our case), but I won’t list them here. Instead, I’ll just print out the variable importance plots for the two measures as these give a good visual overview of the relative importance of variables. The code is a simple one-liner:
The plot is shown in Figure 1 below.
In this case the two measures are pretty consistent so it doesn’t really matter which one you choose.
Random forests are an example of a general class of techniques called ensemble methods. These techniques are based on the principle that averaging over a large number of not-so-good models yields a more reliable prediction than a single model. This is true only if models in the group are independent of each other, which is precisely what bootstrap aggregation and predictor subsetting are intended to achieve.
Although considerably more complex than decision trees, the logic behind random forests is not hard to understand. Indeed, the intuitiveness of the algorithm together with its ease of use and accuracy have made it very popular in the machine learning community.
It is well-known that data-driven stories are a great way to convey results of data science initiatives. What is perhaps not as well-known is that data science projects often have to begin with stories too. Without this “story before the story” there will be no project, no results and no data-driven stories to tell….
For those who prefer to read, here’s a transcript of the video in full:
In the beginning there is no data, let alone results…but there are ideas. So, long before we tell stories about data or results, we have to tell stories about our ideas. The aim of these stories is to get people to care about our ideas as much as we do and, more important, invest in them. Without their interest or investment there will be no results and no further stories to tell.
So one of the first things one has to do is craft a story about the idea…or the story before the story.
Once upon a time there was a CRM system. The system captured every customer interaction that occurred, whether it was by phone, email or face to face conversation. Many quantitative details of interactions were recorded, time, duration, type. And if the interaction led to a sale, the details of the sale were recorded too.
Almost as an aside, the system also gave sales people the opportunity to record their qualitative impressions as free text notes. As you might imagine, this information, though potentially valuable, was never analysed. Sure managers looked at notes in isolation from time to time when referring.to specific customer interactions, but there was no systematic analysis of the corpus as a whole. Nobody had thought it worthwhile to do this, possibly because it is difficult if not quite impossible to analyse unstructured information in the world of relational databases and SQL.
One day, an analyst was browsing data randomly in the system, as good analysts sometimes do. He came across a note that to him seemed like the epitome of a good note…it described what the interaction was about, the customer’s reactions and potential next steps all in a logical fashion.
This gave him an idea. Wouldn’t it be cool, he thought, if we could measure the quality of notes? Not only would this tell us something about the customer and the interaction, it may tell us something about the sales person as well.
The analyst was mega excited…but he realised he’d need help. He was an IT guy and as we all know, business folks in big corporations stopped listening to their IT guys long ago. So our IT guy had his work cut out for him.
After much cogitation, he decided to enlist the help of his friend, a strategic business analyst in the marketing department. This lady, who worked in marketing had the trust of the head of marketing. If she liked the idea, she might be able to help sell it to the head of marketing.
As it turned out, the business analyst loved the idea…more important, since she knew what the sales people do on a day to day basis, she could give the IT guy more ideas on how he could build quantitative measures of the quality of notes. For example, she suggested looking for emotion-laden words or mentions of competitor’s products and so on. The IT guy now had some concrete things to work on. The initial results gave them even more ideas, and soon they had more than enough to make a convincing pitch to the head of marketing.
It would take us too far afield to discuss details of the pitch, but what we will say is this: they avoided technical details, instead focusing on the strategic and innovative aspects of the work.
The marketing head liked the idea…what was there not to like? He agreed to support the effort, and the idea became a project….
…and yes, within months the project resulted in new insights into customer behaviour. But that is another story.
Most techniques of predictive analytics have their origins in probability or statistical theory (see my post on Naïve Bayes, for example). In this post I’ll look at one that has more a commonplace origin: the way in which humans make decisions. When making decisions, we typically identify the options available and then evaluate them based on criteria that are important to us. The intuitive appeal of such a procedure is in no small measure due to the fact that it can be easily explained through a visual. Consider the following graphic, for example:
(Original image: https://www.flickr.com/photos/dullhunk/7214525854, Credit: Duncan Hull)
The tree structure depicted here provides a neat, easy-to-follow description of the issue under consideration and its resolution. The decision procedure is based on asking a series of questions, each of which serve to further reduce the domain of possibilities. The predictive technique I discuss in this post,classification and regression trees (CART), works in much the same fashion. It was invented by Leo Breiman and his colleagues in the 1970s.
In what follows, I will use the open source software, R. If you are new to R, you may want to follow this link for more on the basics of setting up and installing it. Note that the R implementation of the CART algorithm is called RPART (Recursive Partitioning And Regression Trees). This is essentially because Breiman and Co. trademarked the term CART. As some others have pointed out, it is somewhat ironical that the algorithm is now commonly referred to as RPART rather than by the term coined by its inventors.
A bit about the algorithm
The rpart algorithm works by splitting the dataset recursively, which means that the subsets that arise from a split are further split until a predetermined termination criterion is reached. At each step, the split is made based on the independent variable that results in the largest possible reduction in heterogeneity of the dependent (predicted) variable.
Splitting rules can be constructed in many different ways, all of which are based on the notion of impurity- a measure of the degree of heterogeneity of the leaf nodes. Put another way, a leaf node that contains a single class is homogeneous and has impurity=0. There are three popular impurity quantification methods: Entropy (aka information gain), Gini Index and Classification Error. Check out this article for a simple explanation of the three methods.
The rpart algorithm offers the entropy and Gini index methods as choices. There is a fair amount of fact and opinion on the Web about which method is better. Here are some of the better articles I’ve come across:
The answer as to which method is the best is: it depends. Given this, it may be prudent to try out a couple of methods and pick the one that works best for your problem.
Regardless of the method chosen, the splitting rules partition the decision space (a fancy word for the entire dataset) into rectangular regions each of which correspond to a split. Consider the following simple example with two predictors x1 and x2. The first split is at x1=1 (which splits the decision space into two regions x1<1 and x1>1), the second at x2=2, which splits the (x1>1) region into 2 sub-regions, and finally x1=1.5 which splits the (x1>1,x2>2) sub-region further.
It is important to note that the algorithm works by making the best possible choice at each particular stage, without any consideration of whether those choices remain optimal in future stages. That is, the algorithm makes a locally optimal decision at each stage. It is thus quite possible that such a choice at one stage turns out to be sub-optimal in the overall scheme of things. In other words, the algorithm does not find a globally optimal tree.
Another important point relates to well-known bias-variance tradeoff in machine learning, which in simple terms is a tradeoff between the degree to which a model fits the training data and its predictive accuracy. This refers to the general rule that beyond a point, it is counterproductive to improve the fit of a model to the training data as this increases the likelihood of overfitting. It is easy to see that deep trees are more likely to overfit the data than shallow ones. One obvious way to control such overfitting is to construct shallower trees by stopping the algorithm at an appropriate point based on whether a split significantly improves the fit. Another is to grow a tree unrestricted and then prune it back using an appropriate criterion. The rpart algorithm takes the latter approach.
Here is how it works in brief:
Essentially one minimises the cost, , a quantity that is a linear combination of the error (essentially, the fraction of misclassified instances, or variance in the case of a continuous variable), and the number of leaf nodes in the tree, :
First, we note that when , this simply returns the original fully grown tree. As increases, we incur a penalty that is proportional to the number of leaf nodes. This tends to cause the minimum cost to occur for a tree that is a subtree of the original one (since a subtree will have a smaller number of leaf nodes). In practice we vary and pick the value that gives the subtree that results in the smallest cross-validated prediction error. One does not have to worry about programming this because the rpart algorithm actually computes the errors for different values of for us. All we need to do is pick the value of the coefficient that gives the lowest cross-validated error. I will illustrate this in detail in the next section.
An implication of their tendency to overfit data is that decision trees tend to be sensitive to relatively minor changes in the training datasets. Indeed, small differences can lead to radically different looking trees. Pruning addresses this to an extent, but does not resolve it completely. A better resolution is offered by the so-called ensemble methods that average over many differently constructed trees. I’ll discuss one such method at length in a future post.
Finally, I should also mention that decision trees can be used for both classification and regression problems (i.e. those in which the predicted variable is discrete and continuous respectively). I’ll demonstrate both types of problems in the next two sections.
Classification trees using rpart
To demonstrate classification trees, we’ll use the Ionosphere dataset available in the mlbench package in R. I have chosen this dataset because it nicely illustrates the points I wish to make in this post. In general, you will almost always find that algorithms that work fine on classroom datasets do not work so well in the real world…but of course, you know that already!
We begin by setting the working directory, loading the required packages (rpart and mlbench) and then loading the Ionosphere dataset.
Next we separate the data into training and test sets. We’ll use the former to build the model and the latter to test it. To do this, I use a simple scheme wherein I randomly select 80% of the data for the training set and assign the remainder to the test data set. This is easily done in a single R statement that invokes the uniform distribution (runif) and the vectorised function, ifelse. Before invoking runif, I set a seed integer to my favourite integer in order to ensure reproducibility of results.
In the above, I have also removed the training flag from the training and test datasets.
Next we invoke rpart. I strongly recommend you take some time to go through the documentation and understand the parameters and their defaults values. Note that we need to remove the predicted variable from the dataset before passing the latter on to the algorithm, which is why we need to find the column index of the predicted variable (first line below). Also note that we set the method parameter to “class“, which simply tells the algorithm that the predicted variable is discrete. Finally, rpart uses Gini rule for splitting by default, and we’ll stick with this option.
The resulting plot is shown in Figure 3 below. It is quite self-explanatory so I won’t dwell on it here.
Next we see how good the model is by seeing how it fares against the test data.
Note that we need to verify the above results by doing multiple runs, each using different training and test sets. I will do this later, after discussing pruning.
Next, we prune the tree using the cost complexity criterion. Basically, the intent is to see if a shallower subtree can give us comparable results. If so, we’d be better of choosing the shallower tree because it reduces the likelihood of overfitting.
As described earlier, we choose the appropriate pruning parameter (aka cost-complexity parameter) by picking the value that results in the lowest prediction error. Note that all relevant computations have already been carried out by R when we built the original tree (the call to rpart in the code above). All that remains now is to pick the value of :
It is clear from the above, that the lowest cross-validation error (xerror in the table) occurs for (this is CP in the table above). One can find CP programatically like so:
Next, we prune the tree based on this value of CP:
Note that rpart will use a default CP value of 0.01 if you don’t specify one in prune.
The pruned tree is shown in Figure 4 below.
Let’s see how this tree stacks up against the fully grown one shown in Fig 3.
This seems like an improvement over the unpruned tree, but one swallow does not a summer make. We need to check that this holds up for different training and test sets. This is easily done by creating multiple random partitions of the dataset and checking the efficacy of pruning for each. To do this efficiently, I’ll create a function that takes the training fraction, number of runs (partitions) and the name of the dataset as inputs and outputs the proportion of correct predictions for each run. It also optionally prunes the tree. Here’s the code:
Note that in the above, I have set the default value of the prune_tree to FALSE, so the function will execute the first branch of the if statement unless the default is overridden.
OK, so let’s do 50 runs with and without pruning, and check the mean and variance of the results for both sets of runs.
So we see that there is an improvement of about 3% with pruning. Also, if you were to plot the trees as we did earlier, you would see that this improvement is achieved with shallower trees. Again, I point out that this is not always the case. In fact, it often happens that pruning results in worse predictions, albeit with better reliability – a classic illustration of the bias-variance tradeoff.
Regression trees using rpart
In the previous section we saw how one can build decision trees for situations in which the predicted variable is discrete. Let’s now look at the case in which the predicted variable is continuous. We’ll use the Boston Housing dataset from the mlbench package. Much of the discussion of the earlier section applies here, so I’ll just display the code, explaining only the differences.
Next we invoke rpart, noting that the predicted variable is medv (median value of owner-occupied homes in $1000 units) and that we need to set the method parameter to “anova“. The latter tells rpart that the predicted variable is continuous (i.e that this is a regression problem).
The plot of the tree is shown in Figure 5 below.
Next, we need to see how good the predictions are. Since the dependent variable is continuous, we cannot compare the predictions directly against the test set. Instead, we calculate the root mean square (RMS) error. To do this, we request rpart to output the predictions as a vector – one prediction per record in the test dataset. The RMS error can then easily be calculated by comparing this vector with the medv column in the test dataset.
Here is the relevant code:
Again, we need to do multiple runs to check on the reliability of the predictions. However, you already know how to do that so I will leave it to you.
Moving on, we prune the tree using the cost complexity criterion as before. The code is exactly the same as in the classification problem.
The tree is unchanged so I won’t show it here. This means, as far as the cost complexity pruning is concerned, the optimal subtree is the same as the original tree. To confirm this, we’d need to do multiple runs as before – something that I’ve already left as as an exercise for you :). Basically, you’ll need to write a function analogous to the one above, that computes the root mean square error instead of the proportion of correct predictions.
This brings us to the end of my introduction to classification and regression trees using R. Unlike some articles on the topic I have attempted to describe each of the steps in detail and provide at least some kind of a rationale for them. I hope you’ve found the description and code snippets useful.
I’ll end by reiterating a couple points I made early in this piece. The nice thing about decision trees is that they are easy to explain to the users of our predictions. This is primarily because they reflect the way we think about how decisions are made in real life – via a set of binary choices based on appropriate criteria. That said, in many practical situations decision trees turn out to be unstable: small changes in the dataset can lead to wildly different trees. It turns out that this limitation can be addressed by building a variety of trees using different starting points and then averaging over them. This is the domain of the so-called random forest algorithm.We’ll make the journey from decision trees to random forests in a future post.
Postscript, 20th September 2016: I finally got around to finishing my article on random forests.
Graph theory is the an area of mathematics that analyses relationships between pairs of objects. Typically graphs consist of nodes (points representing objects) and edges (lines depicting relationships between objects). As one might imagine, graphs are extremely useful in visualizing relationships between objects. In this post, I provide a detailed introduction to network graphs using R, the premier open source tool statistics package for calculations and the excellent Gephi software for visualization.
The article is organised as follows: I begin by defining the problem and then spend some time developing the concepts used in constructing the graph Following this, I do the data preparation in R and then finally build the network graph using Gephi.
In an introductory article on cluster analysis, I provided an in-depth introduction to a couple of algorithms that can be used to categorise documents automatically. Although these techniques are useful, they do not provide a feel for the relationships between different documents in the collection of interest. In the present piece I show network graphs can be used to to visualise similarity-based relationships within a corpus.
There are many ways to quantify similarity between documents. A popular method is to use the notion of distance between documents. The basic idea is simple: documents that have many words in common are “closer” to each other than those that share fewer words. The problem with distance, however, is that it can be skewed by word count: documents that have an unusually high word count will show up as outliers even though they may be similar (in terms of words used) to other documents in the corpus. For this reason, we will use another related measure of similarity that does not suffer from this problem – more about this in a minute.
Representing documents mathematically
As I explained in my article on cluster analysis, a document can be represented as a point in a conceptual space that has dimensionality equal to the number of distinct words in the collection of documents. I revisit and build on that explanation below.
Say one has a simple document consisting of the words “five plus six”, one can represent it mathematically in a 3 dimensional space in which the individual words are represented by the three axis (See Figure 1). Here each word is a coordinate axis (or dimension). Now, if one connects the point representing the document (point A in the figure) to the origin of the word-space, one has a vector, which in this case is a directed line connecting the point in question to the origin. Specifically, the point A can be represented by the coordinates in this space. This is a nice quantitative representation of the fact that the words five, plus and one appear in the document exactly once. Note, however, that we’ve assumed the order of words does not matter. This is a reasonable assumption in some cases, but not always so.
As another example consider document, B, which consists of only two words: “five plus” (see Fig 2). Clearly this document shares some similarity with document but it is not identical. Indeed, this becomes evident when we note that document (or point) B is simply the point $latex(1, 1, 0)$ in this space, which tells us that it has two coordinates (words/frequencies) in common with document (or point) A.
To be sure, in a realistic collection of documents we would have a large number of distinct words, so we’d have to work in a very high dimensional space. Nevertheless, the same principle holds: every document in the corpus can be represented as a vector consisting of a directed line from the origin to the point to which the document corresponds.
Now it is easy to see that two documents are identical if they correspond to the same point. In other words, if their vectors coincide. On the other hand, if they are completely dissimilar (no words in common), their vectors will be at right angles to each other. What we need, therefore, is a quantity that varies from 0 to 1 depending on whether two documents (vectors) are dissimilar(at right angles to each other) or similar (coincide, or are parallel to each other).
Now here’s the ultra-cool thing, from your high school maths class, you know there is a trigonometric ratio which has exactly this property – the cosine!
What’s even cooler is that the cosine of the angle between two vectors is simply the dot product of the two vectors, which is sum of the products of the individual elements of the vector, divided by the product of the lengths of the two vectors. In three dimensions this can be expressed mathematically as:
where the two vectors are and , and is the angle between the two vectors (see Fig 2).
The upshot of the above is that the cosine of the angle between the vector representation of two documents is a reasonable measure of similarity between them. This quantity, sometimes referred to as cosine similarity, is what we’ll take as our similarity measure in the rest of this article.
The adjacency matrix
If we have a collection of documents, we can calculate the similarity between every pair of documents as we did for A and B in the previous section. This would give us a set of numbers between 0 and 1, which can be conveniently represented as a matrix. This is sometimes called the adjacency matrix. Beware, though, this term has many different meanings in the math literature. I use it in the sense specified above.
Since every document is identical to itself, the diagonal elements of the matrix will all be 1. These similarities are trivial (we know that every document is identical to itself!) so we’ll set the diagonal elements to zero.
Another important practical point is that visualizing every relationship is going to make a very messy graph. There would be edges in such a graph, which would make it impossible to make sense of if we have more than a handful of documents. For this reason, it is normal practice to choose a cutoff value of similarity below which it is set to zero.
Building the adjacency matrix using R
We now have enough background to get down to the main point of this article – visualizing relationships between documents.
The first step is to build the adjacency matrix. In order to do this, we have to build the document term matrix (DTM) for the collection of documents, a process which I have dealt with at length in my introductory pieces on text mining and topic modeling. In fact, the steps are actually identical to those detailed in the second piece. I will therefore avoid lengthy explanations here. However, I’ve listed all the code below with brief comments (for those who are interested in trying this out, the document corpus can be downloaded here and a pdf listing of the R code can be obtained here.)
OK, so here’s the code listing:
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, ““”)
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 rows of a DTM are document vectors akin to the vector representations of documents A and B discussed earlier. The DTM therefore contains all the information we need to calculate the cosine similarity between every pair of documents in the corpus (via equation 1). The R code below implements this, after taking care of a few preliminaries.
A few lines need a brief explanation:
First up, although the DTM is a matrix, it is internally stored in a special form suitable for sparse matrices. We therefore have to explicitly convert it into a proper matrix before using it to calculate similarity.
Second, the names I have given the documents are way too long to use as labels in the network diagram. I have therefore mapped the document names to the row numbers which we’ll use in our network graph later. The mapping back to the original document names is stored in filekey.csv. For future reference, the mapping is shown in Table 1 below.
Table 1: File mappings
Finally, the distance function (as.dist) in the cosine similarity function sets the diagonal elements to zero because the distance between a document and itself is zero…which is just a complicated way of saying that a document is identical to itself 🙂
The last three lines of code above simply implement the cutoff that I mentioned in the previous section. The comments explain the details so I need say no more about it.
…which finally brings us to Gephi.
Visualizing document similarity using Gephi
Gephi is an open source, Java based network analysis and visualisation tool. Before going any further, you may want to download and install it. While you’re at it you may also want to download this excellent quick start tutorial.
Go on, I’ll wait for you…
To begin with, there’s a little formatting quirk that we need to deal with. Gephi expects separators in csv files to be semicolons (;) . So, your first step is to open up the adjacency matrix that you created in the previous section (AdjacencyMatrix.csv) in a text editor and replace commas with semicolons.
Once you’ve done that, fire up Gephi, go to File > Open, navigate to where your Adjacency matrix is stored and load the file. If it loads successfully, you should see a feedback panel as shown in Figure 3. By default Gephi creates a directed graph (i.e one in which the edges have arrows pointing from one node to another). Change this to undirected and click OK.
Once that is done, click on overview (top left of the screen). You should end up with something like Figure 4.
Gephi has sketched out an initial network diagram which depicts the relationships between documents…but it needs a bit of work to make it look nicer and more informative. The quickstart tutorial mentioned earlier describes various features that can be used to manipulate and prettify the graph. In the remainder of this section, I list some that I found useful. Gephi offers many more. Do explore, there’s much more than I can cover in an introductory post.
First some basics. You can:
- Zoom and pan using mouse wheel and right button.
- Adjust edge thicknesses using the slider next to text formatting options on bottom left of main panel.
- Re-center graph via the magnifying glass icon on left of display panel (just above size adjuster).
- Toggle node labels on/off by clicking on grey T symbol on bottom left panel.
Figure 5 shows the state of the diagram after labels have been added and edge thickness adjusted (note that your graph may vary in appearance).
The default layout of the graph is ugly and hard to interpret. Let’s work on fixing it up. To do this, go over to the layout panel on the left. Experiment with different layouts to see what they do. After some messing around, I found the Fruchtermann-Reingold and Force Atlas options to be good for this graph. In the end I used Force Atlas with a Repulsion Strength of 2000 (up from the default of 200) and an Attraction Strength of 1 (down from the default of 10). I also adjusted the figure size and node label font size from the graph panel in the center. The result is shown in Figure 6.
This is much better. For example, it is now evident that document 9 is the most connected one (which table 9 tells us is a transcript of a conversation with Neil Preston on organisational change).
It would be nice if we could colour code edges/nodes and size nodes by their degree of connectivity. This can be done via the ranking panel above the layout area where you’ve just been working.
In the Nodes tab select Degree as the rank parameter (this is the degree of connectivity of the node) and hit apply. Select your preferred colours via the small icon just above the colour slider. Use the colour slider to adjust the degree of connectivity at which colour transitions occur.
Do the same for edges, selecting weight as the rank parameter(this is the degree of similarity between the two douments connected by the edge). With a bit of playing around, I got the graph shown in the screenshot below (Figure 7).
If you want to see numerical values for the rankings, hit the results list icon on the bottom left of the ranking panel. You can see numerical ranking values for both nodes and edges as shown in Figures 8 and 9.
It is easy to see from the figure that documents 21 and 29 are the most similar in terms of cosine ranking. This makes sense, they are pieces in which I have ranted about the current state of enterprise architecture – the first article is about EA in general and the other about the TOGAF framework. If you have a quick skim through, you’ll see that they have a fair bit in common.
Finally, it would be nice if we could adjust node size to reflect the connectedness of the associated document. You can do this via the “gem” symbol on the top right of the ranking panel. Select appropriate min and max sizes (I chose defaults) and hit apply. The node size is now reflective of the connectivity of the node – i.e. the number of other documents to which it is cosine similar to varying degrees. The thickness of the edges reflect the degree of similarity. See Figure 10.
Now that looks good enough to export. To do this, hit the preview tab on main panel and make following adjustments to the default settings:
Under Node Labels:
1. Check Show Labels
2. Uncheck proportional size
3. Adjust font to required size
1. Change thickness to 10
2. Check rescale weight
Hit refresh after making the above adjustments. You should get something like Fig 11.
All that remains now is to do the deed: hit export SVG/PDF/PNG to export the diagram. My output is displayed in Figure 12. It clearly shows the relationships between the different documents (nodes) in the corpus. The nodes with the highest connectivity are indicated via node size and colour (purple for high, green for low) and strength of similarity is indicated by edge thickness.
…which brings us to the end of this journey.
The techniques of text analysis enable us to quantify relationships between documents. Document similarity is one such relationship. Numerical measures are good, but the comprehensibility of these can be further enhanced through meaningful visualisations. Indeed, although my stated objective in this article was to provide an introduction to creating network graphs using Gephi and R (which I hope I’ve succeeded in doing), a secondary aim was to show how document similarity can be quantified and visualised. I sincerely hope you’ve found the discussion interesting and useful.
Many thanks for reading! As always, your feedback would be greatly appreciated.
One of the key problems of predictive analytics is to classify entities or events based on a knowledge of their attributes. An example: one might want to classify customers into two categories, say, ‘High Value’ or ‘Low Value,’ based on a knowledge of their buying patterns. Another example: to figure out the party allegiances of representatives based on their voting records. And yet another: to predict the species a particular plant or animal specimen based on a list of its characteristics. Incidentally, if you haven’t been there already, it is worth having a look at Kaggle to get an idea of some of the real world classification problems that people tackle using techniques of predictive analytics.
Given the importance of classification-related problems, it is no surprise that analytics tools offer a range of options. My favourite (free!) tool, R, is no exception: it has a plethora of state of the art packages designed to handle a wide range of problems. One of the problems with this diversity of choice is that it is often confusing for beginners to figure out which one to use in a particular situation. Over the next several months, I intend to write up tutorial articles covering many of the common algorithms, with a particular focus on their strengths and weaknesses; explaining where they work well and where they don’t. I’ll kick-off this undertaking with a simple yet surprisingly effective algorithm – the Naïve Bayes classifier.
Just enough theory
I’m going to assume you have R and RStudio installed on your computer. If you need help with this, please follow the instructions here.
To introduce the Naive Bayes algorithm, I will use the HouseVotes84 dataset, which contains US congressional voting records for 1984. The data set is in the mlbench package which is not part of the base R installation. You will therefore need to install it if you don’t have it already. Package installation is a breeze in RStudio – just go to Tools > Install Packages and follow the prompts.
The HouseVotes84 dataset describes how 435 representatives voted – yes (y), no (n) or unknown (NA) – on 16 key issues presented to Congress. The dataset also provides the party affiliation of each representative – democrat or republican.
Let’s begin by exploring the dataset. To do this, we load mlbench, fetch the dataset and get some summary stats on it. (Note: a complete listing of the code in this article can be found here)
It is good to begin by exploring the data visually. To this end, let’s do some bar plots using the basic graphic capabilities of R:
The plots are shown in Figures 1 through 3.
Among other things, such plots give us a feel for the probabilities associated with how representatives from parties tend to vote on specific issues.
The classification problem at hand is to figure out the party affiliation from a knowledge of voting patterns. For simplicity let us assume that there are only 3 issues voted on instead of the 16 in the actual dataset. In concrete terms we wish to answer the question, “what is the probability that a representative is, say, a democrat (D) given that he or she has voted, say, on the three issues?” To keep things simple I’m assuming there are no NA values.
In the notation of conditional probability this can be written as,
(Note: If you need a refresher on conditional probability, check out this post for a simple explanation.)
We’re interested only in relative probabilities of the representative being a democrat or republican because the predicted party affiliation depends only on which of the two probabilities is larger (the actual value of the probability is not important). This being the case, we can factor out any terms that are constant. As it happens, the denominator of the above equation – the probability of a particular voting pattern – is a constant because it depends on the total number of representatives (from both parties) who voted a particular way.
Now, using the chain rule of conditional probability, we can rewrite the numerator as:
Basically, the second term on the left hand side, , is the probability of getting a particular voting pattern (y,n,y) assuming the rep is a Democrat (D). The definition of conditional probability allows us to rewrite this as the probability of getting a n vote for issue v2 and a y vote for issue v3 given that the rep is a Democrat who has voted y on issue v1. Again, this is simply a consequence of the definition of conditional probability.
Another application of the chain rule gives:
Where we have now factored out the n vote on the second issue.
The key assumption of Naïve Bayes is that the conditional probability of each feature given the class is independent of all other features. In mathematical terms this means that,
The quantity of interest, the numerator of equation (1) can then be written as:
The assumption of independent conditional probabilities is a drastic one. What it is saying is that the features are completely independent of each other. This is clearly not the case in the situation above: how representatives vote on a particular issue is coloured by their beliefs and values. For example, the conditional probability of voting patterns on socially progressive issues are definitely not independent of each other. However, as we shall see in the next section, the Naïve Bayes assumption works well for this problem as it does in many other situations where we know upfront that it is grossly incorrect.
Another good example of the unreasonable efficacy of Naive Bayes is in spam filtering. In the case of spam, the features are individual words in an email. It is clear that certain word combinations tend to show up consistently in spam – for example, “online”, “meds”, “Viagra” and “pharmacy.” In other words, we know upfront that their occurrences are definitely not independent of each other. Nevertheless, Naïve Bayes based spam detectors which assume mutual independence of features do remarkably well in distinguishing spam from ham.
Why is this so?
To explain why, I return to a point I mentioned earlier: to figure out the affiliation associated with a particular voting pattern (say, v1=y, v2=n,v3=y) one only needs to know which of the two probabilities and is greater than the other. That is, the values of these probabilities are not important in determining the party affiliations.
This hints as to why the independence assumption might not be so quite so idiotic. Since the prediction depends only the on the maximum, the algorithm will get it right even if there are dependencies between feature providing the dependencies do not change which class has the maximum probability (once again, note that only the maximal class is important here, not the value of the maximum).
Yet another reason for the surprising success of Naïve Bayes is that dependencies often cancel out across a large set of features. But, of course, there is no guarantee that this will always happen.
In general, Naïve Bayes algorithms work better for problems in which the dependent (predicted) variable is discrete, even when there are dependencies between features (spam detection is a good example). They work less well for regression problems – i.e those in which predicted variables are continuous.
I hope the above has given you an intuitive feel for how Naïve Bayes algorithms work. I don’t know about you, but my head’s definitely spinning after writing out all that mathematical notation.
It’s time to clear our heads by doing some computation.
Naïve Bayes in action
There are a couple of well-known implementations of Naïve Bayes in R. One of them is the naiveBayes method in the e1071 package and the other is NaiveBayes method in the klaR package. I’ll use the former for no other reason than it seems to be more popular. That said, I have used the latter too and can confirm that it works just as well.
We’ve already loaded and explored the HouseVotes84 dataset. One of the things you may have noticed when summarising the data is that there are a fair number of NA values. Naïve Bayes algorithms typically handle NA values either by ignoring records that contain any NA values or by ignoring just the NA values. These choices are indicated by the value of the variable na.action in the naiveBayes algorithm, which is set to na.omit (to ignore the record) or na.pass (to ignore the value).
Just for fun, we’ll take a different approach. We’ll impute NA values for a given issue and party by looking at how other representatives from the same party voted on the issue. This is very much in keeping with the Bayesian spirit: we infer unknowns based on a justifiable belief – that is, belief based on the evidence.
To do this I write two functions: one to compute the number of NA values for a given issue (vote) and class (party affiliation), and the other to calculate the fraction of yes votes for a given issue (column) and class (party affiliation).
sum_y<-sum(HouseVotes84[,col]==’y’ & HouseVotes84$Class==cls,na.rm = TRUE)
sum_n<-sum(HouseVotes84[,col]==’n’ & HouseVotes84$Class==cls,na.rm = TRUE)
Before proceeding, you might want to go back to the data and convince yourself that these values are sensible.
We can now impute the NA values based on the above. We do this by randomly assigning values ( y or n) to NAs, based on the proportion of members of a party who have voted y or n. In practice, we do this by invoking the uniform distribution and setting an NA value to y if the random number returned is less than the probability of a yes vote and to n otherwise. This is not as complicated as it sounds; you should be able to figure the logic out from the code below.
c1 <- which(is.na(HouseVotes84[,i])& HouseVotes84$Class==’democrat’,arr.ind = TRUE)
c2 <- which(is.na(HouseVotes84[,i])& HouseVotes84$Class==’republican’,arr.ind = TRUE)
Note that the which function filters indices by the criteria specified in the arguments and ifelse is a vectorised conditional function which enables us to apply logical criteria to multiple elements of a vector.
At this point it is a good idea to check that the NAs in each column have been set according to the voting patterns of non-NAs for a given party. You can use the p_y_col_class() function to check that the new probabilities are close to the old ones. You might want to do this before you proceed any further.
The next step is to divide the available data into training and test datasets. The former will be used to train the algorithm and produce a predictive model. The effectiveness of the model will then be tested using the test dataset. There is a great deal of science and art behind the creation of training and testing datasets. An important consideration is that both sets must contain records that are representative of the entire dataset. This can be difficult to do, especially when data is scarce and there are predictors that do not vary too much…or vary wildly for that matter. On the other hand, problems can also arise when there are redundant predictors. Indeed, the much of the art of successful prediction lies in figuring out which predictors are likely to lead to better predictions, an area known as feature selection. However, that’s a topic for another time. Our current dataset does not suffer from any of these complications so we’ll simply divide the it in an 80/20 proportion, assigning the larger number of records to the training set.
Now we’re finally good to build our Naive Bayes model (machine learning folks call this model training rather than model building – and I have to admit, it does sound a lot cooler).
The code to train the model is anticlimactically simple:
Here we’ve invokedthe naiveBayes method from the e1071 package. The first argument uses R’s formula notation.In this notation, the dependent variable (to be predicted) appears on the left hand side of the ~ and the independent variables (predictors or features) are on the right hand side. The dot (.) is simply shorthand for “all variable other than the dependent one.” The second argument is the dataframe that contains the training data. Check out the documentation for the other arguments of naiveBayes; it will take me too far afield to cover them here. Incidentally, you can take a look at the model using the summary() or str() functions, or even just entering the model name in the R console:
Note that I’ve suppressed the output above.
Now that we have a model, we can do some predicting. We do this by feeding our test data into our model and comparing the predicted party affiliations with the known ones. The latter is done via the wonderfully named confusion matrix – a table in which true and predicted values for each of the predicted classes are displayed in a matrix format. This again is just a couple of lines of code:
The numbers you get will be different because your training/test sets are almost certainly different from mine.
In the confusion matrix (as defined above), the true values are in columns and the predicted values in rows. So, the algorithm has correctly classified 38 out of 43 (i.e. 38+5) Democrats and 22 out of 25 Republicans (i.e. 22+3). That’s pretty decent. However, we need to keep in mind that this could well be quirk of the choice of dataset. To address this, we should get a numerical measure of the efficacy of the algorithm and for different training and testing datasets. A simple measure of efficacy would be the fraction of predictions that the algorithm gets right. For the training/testing set above, this is simply 60/68 (see the confusion matrix above). The simplest way to calculate this in R is:
A natural question to ask at this point is: how good is this prediction. This question cannot be answered with only a single run of the model; we need to do many runs and look at the spread of the results. To do this, we’ll create a function which takes the number of times the model should be run and the training fraction as inputs and spits out a vector containing the proportion of correct predictions for each run. Here’s the function
I’ve not commented the above code as it is essentially a repeat of the steps described earlier. Also, note that I have not made any effort to make the code generic or efficient.
Let’s do 20 runs with the same training fraction (0.8) as before:
 0.9102564 0.9080460 0.9139785 0.9200000 0.9090909 0.9239130 0.9605263 0.9333333
 0.9052632 0.8977273 0.9642857 0.8518519
0.8519 0.9074 0.9170 0.9177 0.9310 0.9643
We see that the outcome of the runs are quite close together, in the 0.85 to 0.95 range with a standard deviation of 0.025. This tells us that Naive Bayes does a pretty decent job with this data.
I originally intended to cover a few more case studies in this post, a couple of which highlight the shortcomings of the Naive Bayes algorithm. However, I realize that doing so would make this post unreasonably long, so I’ll stop here with a few closing remarks, and a promise to write up the rest of the story in a subsequent post.
To sum up: I have illustrated the use of a popular Naive Bayes implementation in R and attempted to convey an intuition for how the algorithm works. As we have seen, the algorithm works quite well in the example case, despite the violation of the assumption of independent conditional probabilities.
The reason for the unreasonable effectiveness of the algorithm is two-fold. Firstly, the algorithm picks the predicted class based on the largest predicted probability, so ordering is more important than the actual value of the probability. Secondly, in many cases, a bias one way for a particular vote may well be counteracted by a bias the other way for another vote. That is, biases tend to cancel out, particularly if there are a large number of features.
That said, there are many cases in which the algorithm fails miserably – and we’ll look at some of these in a future post. However, despite its well known shortcomings, Naive Bayes is often the first port of call in prediction problems simply because it is easy to set up and is fast compared to many of the iterative algorithms we will explore later in this series of articles.
Thanks for reading! If you liked this piece, you might enjoy the other articles in my “Gentle introduction to analytics using R” series. Here are the links: