Community Detection in Graphs — a Casual Tour

Graphs are among the most interesting and useful objects in mathematics. Any situation or idea that can be described by objects with connections is a graph, and one of the most prominent examples of a real-world graph that one can come up with is a social network.

Recall, if you aren’t already familiar with this blog’s gentle introduction to graphs, that a graph G is defined by a set of vertices V, and a set of edges E, each of which connects two vertices. For this post the edges will be undirected, meaning connections between vertices are symmetric.

One of the most common topics to talk about for graphs is the notion of a community. But what does one actually mean by that word? It’s easy to give an informal definition: a subset of vertices C such that there are many more edges between vertices in C than from vertices in C to vertices in V - C (the complement of C). Try to make this notion precise, however, and you open a door to a world of difficult problems and open research questions. Indeed, nobody has yet come to a conclusive and useful definition of what it means to be a community. In this post we’ll see why this is such a hard problem, and we’ll see that it mostly has to do with the word “useful.” In future posts we plan to cover some techniques that have found widespread success in practice, but this post is intended to impress upon the reader how difficult the problem is.

The simplest idea

The simplest thing to do is to say a community is a subset of vertices which are completely connected to each other. In the technical parlance, a community is a subgraph which forms a clique. Sometimes an n-clique is also called a complete graph on n vertices, denoted K_n. Here’s an example of a 5-clique in a larger graph:

"Where's Waldo" for graph theorists: a clique hidden in a larger graph.

“Where’s Waldo” for graph theorists: a clique hidden in a larger graph.

Indeed, it seems reasonable that if we can reliably find communities at all, then we should be able to find cliques. But as fate should have it, this problem is known to be computationally intractable. In more detail, the problem of finding the largest clique in a graph is NP-hard. That essentially means we don’t have any better algorithms to find cliques in general graphs than to try all possible subsets of the vertices and check to see which, if any, form cliques. In fact it’s much worse, this problem is known to be hard to approximate to any reasonable factor in the worst case (the error of the approximation grows polynomially with the size of the graph!). So we can’t even hope to find a clique half the size of the biggest, or a thousandth the size!

But we have to take these impossibility results with a grain of salt: they only say things about the worst case graphs. And when we’re looking for communities in the real world, the worst case will never show up. Really, it won’t! In these proofs, “worst case” means that they encode some arbitrarily convoluted logic problem into a graph, so that finding the clique means solving the logic problem. To think that someone could engineer their social network to encode difficult logic problems is ridiculous.

So what about an “average case” graph? To formulate this typically means we need to consider graphs randomly drawn from a distribution.

Random graphs

The simplest kind of “randomized” graph you could have is the following. You fix some set of vertices, and then run an experiment: for each pair of vertices you flip a coin, and if the coin is heads you place an edge and otherwise you don’t. This defines a distribution on graphs called G(n, 1/2), which we can generalize to G(n, p) for a coin with bias p. With a slight abuse of notation, we call G(n, p) the Erdős–Rényi random graph (it’s not a graph but a distribution on graphs). We explored this topic form a more mathematical perspective earlier on this blog.

So we can sample from this distribution and ask questions like: what’s the probability of the largest clique being size at least 20? Indeed, cliques in Erdős–Rényi random graphs are so well understood that we know exactly how they work. For example, if p=1/2 then the size of the largest clique is guaranteed (with overwhelming probability as n grows) to have size k(n) or k(n)+1, where k(n) is about 2 \log n. Just as much is known about other values of p as well as other properties of G(n,p), see Wikipedia for a short list.

In other words, if we wanted to find the largest clique in an Erdős–Rényi random graph, we could check all subsets of size roughly 2\log(n), which would take about (n / \log(n))^{\log(n)} time. This is pretty terrible, and I’ve never heard of an algorithm that does better (contrary to the original statement in this paragraph that showed I can’t count). In any case, it turns out that the Erdős–Rényi random graph, and using cliques to represent communities, is far from realistic. There are many reasons why this is the case, but here’s one example that fits with the topic at hand. If I thought the world’s social network was distributed according to G(n, 1/2) and communities were cliques, then I would be claiming that the largest community is of size 65 or 66. Estimated world population: 7 billion, 2 \log(7 \cdot 10^9) \sim 65. Clearly this is ridiculous: there are groups of larger than 66 people that we would want to call “communities,” and there are plenty of communities that don’t form bona-fide cliques.

Another avenue shows that things are still not as easy as they seem in Erdős–Rényi land. This is the so-called planted clique problem. That is, you draw a graph G from G(n, 1/2). You give G to me and I pick a random but secret subset of r vertices and I add enough edges to make those vertices form an r-clique. Then I ask you to find the r-clique. Clearly it doesn’t make sense when r < 2 \log (n) because you won’t be able to tell it apart from the guaranteed cliques in G. But even worse, nobody knows how to find the planted clique when r is even a little bit smaller than \sqrt{n} (like, r = n^{9/20} even). Just to solidify this with some numbers, we don’t know how to reliably find a planted clique of size 60 in a random graph on ten thousand vertices, but we do when the size of the clique goes up to 100. The best algorithms we know rely on some sophisticated tools in spectral graph theory, and their details are beyond the scope of this post.

So Erdős–Rényi graphs seem to have no hope. What’s next? There are a couple of routes we can take from here. We can try to change our random graph model to be more realistic. We can relax our notion of communities from cliques to something else. We can do both, or we can do something completely different.

Other kinds of random graphs

There is an interesting model of Barabási and Albert, often called the “preferential attachment” model, that has been described as a good model of large, quickly growing networks like the internet. Here’s the idea: you start off with a two-clique G = K_2, and at each time step t you add a new vertex v to G, and new edges so that the probability that the edge (v,w) is added to G is proportional to the degree of w (as a fraction of the total number of edges in G). Here’s an animation of this process:

Image source: Wikipedia

The significance of this random model is that it creates graphs with a small number of hubs, and a large number of low-degree vertices. In other words, the preferential attachment model tends to “make the rich richer.” Another perspective is that the degree distribution of such a graph is guaranteed to fit a so-called power-law distribution. Informally, this means that the overall fraction of small-degree vertices gives a significant contribution to the total number of edges. This is sometimes called a “fat-tailed” distribution. Since power-law distributions are observed in a wide variety of natural settings, some have used this as justification for working in the preferential attachment setting. On the other hand, this model is known to have no significant community structure (by any reasonable definition, certainly not having cliques of nontrivial size), and this has been used as evidence against the model. I am not aware of any work done on planting dense subgraphs in graphs drawn from a preferential attachment model, but I think it’s likely to be trivial and uninteresting. On the other hand, Bubeck et al. have looked at changing the initial graph (the “seed”) from a 2-clique to something else, and seeing how that affects the overall limiting distribution.

Another model that often shows up is a model that allows one to make a random graph starting with any fixed degree distribution, not just a power law. There are a number of models that do this to some fashion, and you’ll hear a lot of hyphenated names thrown around like Chung-Lu and Molloy-Reed and Newman-Strogatz-Watts. The one we’ll describe is quite simple. Say you start with a set of vertices V, and a number d_v for each vertex v, such that the sum of all the d_v is even. This condition is required because in any graph the sum of the degrees of a vertex is twice the number of edges. Then you imagine each vertex v having d_v “edge-stubs.” The name suggests a picture like the one below:

Each node has a prescribed number of "edge stubs," which are randomly connected to form a graph.

Each node has a prescribed number of “edge stubs,” which are randomly connected to form a graph.

Now you pick two edge stubs at random and connect them. One usually allows self-loops and multiple edges between vertices, so that it’s okay to pick two edge stubs from the same vertex. You keep doing this until all the edge stubs are accounted for, and this is your random graph. The degrees were fixed at the beginning, so the only randomization is in which vertices are adjacent. The same obvious biases apply, that any given vertex is more likely to be adjacent to high-degree vertices, but now we get to control the biases with much more precision.

The reason such a model is useful is that when you’re working with graphs in the real world, you usually have statistical information available. It’s simple to compute the degree of each vertex, and so you can use this random graph as a sort of “prior” distribution and look for anomalies. In particular, this is precisely how one of the leading measures of community structure works: the measure of modularity. We’ll talk about this in the next section.

Other kinds of communities

Here’s one easy way to relax our notion of communities. Rather than finding complete subgraphs, we could ask about finding very dense subgraphs (ignoring what happens outside the subgraph). We compute density as the average degree of vertices in the subgraph.

If we impose no bound on the size of the subgraph an algorithm is allowed to output, then there is an efficient algorithm for finding the densest subgraph in a given graph. The general exact solution involves solving a linear programming problem and a little extra work, but luckily there is a greedy algorithm that can get within half of the optimal density. You start with all the vertices S_n = V, and remove any vertex of minimal degree to get S_{n-1}. Continue until S_0, and then compute the density of all the S_i. The best one is guaranteed to be at least half of the optimal density. See this paper of Moses Charikar for a more formal analysis.

One problem with this is that the size of the densest subgraph might be too big. Unfortunately, if you fix the size of the dense subgraph you’re looking for (say, you want to find the densest subgraph of size at most k where k is an input), then the problem once again becomes NP-hard and suffers from the same sort of inapproximability theorems as finding the largest clique.

A more important issue with this is that a dense subgraph isn’t necessarily a community. In particular, we want communities to be dense on the inside and sparse on the outside. The densest subgraph analysis, however, might rate the following graph as one big dense subgraph instead of two separately dense communities with some modest (but not too modest) amount of connections between them.

What should be identified as communities?

What are the correct communities here?

Indeed, we want a quantifiable a notion of “dense on the inside and sparse on the outside.” One such formalization is called modularity. Modularity works as follows. If you give me some partition of the vertices of G into two sets, modularity measures how well this partition reflects two separate communities. It’s the definition of “community” here that makes it interesting. Rather than ask about densities exactly, you can compare the densities to the expected densities in a given random graph model.

In particular, we can use the fixed-degree distribution model from the last section. If we know the degrees of all the vertices ahead of time, we can compute the probability that we see some number of edges going between the two pieces of the partition relative to what we would see at random. If the difference is large (and largely biased toward fewer edges across the partition and more edges within the two subsets), then we say it has high modularity. This involves a lot of computations  — the whole measure can be written as a quadratic form via one big matrix — but the idea is simple enough. We intend to write more about modularity and implement the algorithm on this blog, but the excited reader can see the original paper of M.E.J. Newman.

Now modularity is very popular but it too has shortcomings. First, even though you can compute the modularity of a given partition, there’s still the problem of finding the partition that globally maximizes modularity. Sadly, this is known to be NP-hard. Mover, it’s known to be NP-hard even if you’re just trying to find a partition into two pieces that maximizes modularity, and even still when the graph is regular (every vertex has the same degree).

Still worse, while there are some readily accepted heuristics that often “do well enough” in practice, we don’t even know how to approximate modularity very well. Bhaskar DasGupta has a line of work studying approximations of maximum modularity, and he has proved that for dense graphs you can’t even approximate modularity to within any constant factor. That is, the best you can do is have an approximation that gets worse as the size of the graph grows. It’s similar to the bad news we had for finding the largest clique, but not as bad. For example, when the graph is sparse it’s known that one can approximate modularity to within a \log(n) factor of the optimum, where n is the number of vertices of the graph (for cliques the factor was like n^c for some c, and this is drastically worse).

Another empirical issue is that modularity seems to fail to find small communities. That is, if your graph has some large communities and some small communities, strictly maximizing the modularity is not the right thing to do. So we’ve seen that even the leading method in the field has some issues.

Something completely different

The last method I want to sketch is in the realm of “something completely different.” The notion is that if we’re given a graph, we can run some experiment on the graph, and the results of that experiment can give us insight into where the communities are.

The experiment I’m going to talk about is the random walk. That is, say you have a vertex v in a graph G and you want to find some vertices that are “closest” to v. That is, those that are most likely to be in the same community as v. What you can do is run a random walk starting at v. By a “random walk” I mean you start at v, you pick a neighbor at random and move to it, then repeat. You can compute statistics about the vertices you visit in a sample of such walks, and the vertices that you visit most often are those you say are “in the same community as v. One important parameter is how long the walk is, but it’s generally believed to be best if you keep it between 3-6 steps.

Of course, this is not a partition of the vertices, so it’s not a community detection algorithm, but you can turn it into one. Run this process for each vertex, and use it to compute a “distance” between all the pairs of vertices. Then you compute a tree of partitions by lumping the closest pairs of vertices into the same community, one at a time, until you’ve got every vertex. At each step of the way, you compute the modularity of the partition, and when you’re done you choose the partition that maximizes modularity. This algorithm as a whole is called the walktrap clustering algorithm, and was introduced by Pons and Latapy in 2005.

This sounds like a really great idea, because it’s intuitive: there’s a relatively high chance that the friends of your friends are also your friends. It’s also really great because there is an easily measurable tradeoff between runtime and quality: you can tune down the length of the random walk, and the number of samples you take for each vertex, to speed up the runtime but lower the quality of your statistical estimates. So if you’re working on huge graphs, you get a lot of control and a clear idea of exactly what’s going on inside the algorithm (something which is not immediately clear in a lot of these papers).

Unfortunately, I’m not aware of any concrete theoretical guarantees for walktrap clustering. The one bit of theoretical justification I’ve read over the last year is that you can relate the expected distances you get to certain spectral properties of the graph that are known to be related to community structure, but the lower bounds on maximizing modularity already suggest (though they do not imply) that walktrap won’t do that well in the worst case.

So many algorithms, so little time!

I have only brushed the surface of the literature on community detection, and the things I have discussed are heavily biased toward what I’ve read about and used in my own research. There are methods based on information theory, label propagation, and obscure physics processes like “spin glass” (whatever that is, it sounds frustrating).

And we have only been talking about perfect community structure. What if you want to allow people to be in multiple communities, or have communities at varying levels of granularity (e.g. a sports club within a school versus the whole student body of that school)? What if we want to allow people to be “members” of a community at varying degrees of intensity? How do we deal with noisy signals in our graphs? For example, if we get our data from observing people talk, are two people who have heated arguments considered to be in the same community? Since a lot social network data comes from sources like Twitter and Facebook where arguments are rampant, how do we distinguish between useful and useless data? More subtly, how do we determine useful information if a group within the social network are trying to mask their discovery? That is, how do we deal with adversarial noise in a graph?

And all of this is just on static graphs! What about graphs that change over time? You can keep making the problem more and more complicated as it gets more realistic.

With the huge wealth of research that has already been done just on the simplest case, and the difficult problems and known barriers to success even for the simple problems, it seems almost intimidating to even begin to try to answer these questions. But maybe that’s what makes them fascinating, not to mention that governments and big businesses pour many millions of dollars into this kind of research.

In the future of this blog we plan to derive and implement some of the basic methods of community detection. This includes, as a first outline, the modularity measure and the walktrap clustering algorithm. Considering that I’m also going to spend a large part of the summer thinking about these problems (indeed, with some of the leading researchers and upcoming stars under the sponsorship of the American Mathematical Society), it’s unlikely to end there.

Until next time!

About these ads

k-Means Clustering and Birth Rates

A common problem in machine learning is to take some kind of data and break it up into “clumps” that best reflect how the data is structured. A set of points which are all collectively close to each other should be in the same clump.

A simple picture will clarify any vagueness in this:

cluster-example

Here the data consists of points in the plane. There is an obvious clumping of the data into three pieces, and we want a way to automatically determine which points are in which clumps. The formal name for this problem is the clustering problem. That is, these clumps of points are called clusters, and there are various algorithms which find a “best” way to split the data into appropriate clusters.

The important applications of this are inherently similarity-based: if our data comes from, say, the shopping habits of users of some website, we may want to target a group of shoppers who buy similar products at similar times, and provide them with a coupon for a specific product which is valid during their usual shopping times. Determining exactly who is in that group of shoppers (and more generally, how many groups there are, and what the features the groups correspond to) if the main application of clustering.

This is something one can do quite easily as a human on small visualizable datasets, but the usual the digital representation (a list of numeric points with some number of dimensions) doesn’t yield any obvious insights. Moreover, as the data becomes more complicated (be it by dimension increase, data collection errors, or sheer volume) the “human method” can easily fail or become inconsistent. And so we turn to mathematics to formalize the question.

In this post we will derive one possible version of the clustering problem known as the k-means clustering or centroid clustering problem, see that it is a difficult problem to solve exactly, and implement a heuristic algorithm in place of an exact solution.

And as usual, all of the code used in this post is freely available on this blog’s Github page.

Partitions and Squared Deviations

The process of clustering is really a process of choosing a good partition of the data. Let’s call our data set S, and formalize it as a list of points in space. To be completely transparent and mathematical, we let S be a finite subset of a metric space (X,d), where d is our distance metric.

Definition: We call a partition of a set S a choice of subsets A_1, \dots, A_n of S so that every element of S is in exactly one of the A_i.

A couple of important things to note about partitions is that the union of all the A_i is S, and that any two A_i, A_j intersect trivially. These are immediate consequences of the definition, and together provide an equivalent, alternative definition for a partition. As a simple example, the even and odd integers form a partition of the whole set of integers.

There are many different kinds of clustering problems, but every clustering problem seeks to partition a data set in some way depending on the precise formalization of the goal of the problem. We should note that while this section does give one of many possible versions of this problem, it culminates in the fact that this formalization is too hard to solve exactly. An impatient reader can safely skip to the following section where we discuss the primary heuristic algorithm used in place of an exact solution.

In order to properly define the clustering problem, we need to specify the desired features of a cluster, or a desired feature of the set of all clusters combined. Intuitively, we think of a cluster as a bunch of points which are all close to each other. We can measure this explicitly as follows. Let A be a fixed subset of the partition we’re interested in. Then we might want to optimize the sum of all of the distances of pairs of points within A to be a measure of it’s “clusterity.” In symbols, this would be

\displaystyle \sum_{x \neq y \in A} d(x, y)

If this quantity is small, then it says that all of the points in the cluster A are close to each other, and A is a good cluster. Of course, we want all clusters to be “good” simultaneously, so we’d want to minimize the sum of these sums over all subsets in the partition.

Note that if there are n points in A, then the above sum involves \choose{n}{2} \sim n^2 distance calculations, and so this could get quite inefficient with large data sets. One of the many alternatives is to pick a “center” for each of the clusters, and try to minimize the sum of the distances of each point in a cluster from its center. Using the same notation as above, this would be

\displaystyle \sum_{x \in A} d(x, c)

where c denotes the center of the cluster A. This only involves n distance calculations, and is perhaps a better measure of “clusterity.” Specifically, if we use the first option and one point in the cluster is far away from the others, we essentially record that single piece of information n - 1 times, whereas in the second we only record it once.

The method we will use to determine the center can be very general. We could use one of a variety of measures of center, like the arithmetic mean, or we could try to force one of the points in A to be considered the “center.” Fortunately, the arithmetic mean has the property that it minimizes the above sum for all possible choices of c. So we’ll stick with that for now.

And so the clustering problem is formalized.

Definition: Let (X,d) be a metric space with metric d, and let S \subset (X,d) be a finite subset. The centroid clustering problem is the problem of finding for any positive integer k a partition \left \{ A_1 ,\dots A_k \right \} of S so that the following quantity is minimized:

\displaystyle \sum_{i=1}^k\sum_{x \in A_i} d(x, c(A_i))

where c(A_i) denotes the center of a cluster, defined as the arithmetic mean of the points in A_i:

\displaystyle c(A) = \frac{1}{|A|} \sum_{x \in A} x

Before we continue, we have a confession to make: the centroid clustering problem is prohibitively difficult. In particular, it falls into a class of problems known as NP-hard problems. For the working programmer, NP-hard means that there is unlikely to be an exact solution to the problem which is better than trying all possible partitions.

We’ll touch more on this after we see some code, but the salient fact is that a heuristic algorithm is our best bet. That is, all of this preparation with partitions and squared deviations really won’t come into the algorithm design at all. Formalizing this particular problem in terms of sets and a function we want to optimize only allows us to rigorously prove it is difficult to solve exactly. And so, of course, we will develop a naive and intuitive heuristic algorithm to substitute for an exact solution, observing its quality in practice.

Lloyd’s Algorithm

The most common heuristic for the centroid clustering problem is Lloyd’s algorithm, more commonly known as the k-means clustering algorithm. It was named after its inventor Stuart Lloyd, a University of Chicago graduate and member of the Manhattan project who designed the algorithm in 1957 during his time at Bell Labs.

Heuristics tend to be on the simpler side, and Lloyd’s algorithm is no exception. We start by fixing a number of clusters k and choosing an arbitrary initial partition A = \left \{ A_1, \dots, A_k \right \}. The algorithm then proceeds as follows:

repeat:
   compute the arithmetic mean c[i] of each A[i]
   construct a new partition B:
      each subset B[i] is given a center c[i] computed from A
      x is assigned to the subset B[i] whose c[i] is closest
   stop if B is equal to the old partition A, else set A = B

Intuitively, we imagine the centers of the partitions being pulled toward the center of mass of the points in its currently assigned cluster, and then the points deciding selectively who to pull towards them. (Indeed, precisely because of this the algorithm may not always give sensible results, but more on that later.)

One who is in tune with their inner pseudocode will readily understand the above algorithm. But perhaps the simplest way to think about this algorithm is functionally. That is, we are constructing this partition-updating function f which accepts as input a partition of the data and produces as output a new partition as follows: first compute the mean of centers of the subsets in the old partition, and then create the new partition by gathering all the points closest to each center. These are the fourth and fifth lines of the pseudocode above.

Indeed, the rest of the pseudocode is merely pomp and scaffolding! What we are really after is a fixed point of the partition-updating function f. In other words, we want a partition P such that f(P) = P. We go about finding one in this algorithm by applying f to our initial partition A, and then recursively applying f to its own output until we no longer see a change.

Perhaps we should break away from traditional pseudocode illegibility and rewrite the algorithm as follows:

define updatePartition(A):
   let c[i] = center(A[i])
   return a new partition B:
      each B[i] is given the points which are closest to c[i]

compute a fixed point by recursively applying 
updatePartition to any initial partition.

Of course, the difference between these pseudocode snippets is just the difference between functional and imperative programming. Neither is superior, but the perspective of both is valuable in its own right.

And so we might as well implement Lloyd’s algorithm in two such languages! The first, weighing in at a whopping four lines, is our Mathematica implementation:

closest[x_, means_] :=
  means[[First[Ordering[Map[EuclideanDistance[x, #] &, means]]]]];

partition[points_, means_] := GatherBy[points, closest[#, means]&];
updatePartition[points_, old_] := partition[points, Map[Mean, old]];

kMeans[points_, initialMeans_] := FixedPoint[updatePartition[points, #]&, partition[points, initialMeans]];

While it’s a little bit messy (as nesting 5 function calls and currying by hand will inevitably be), the ideas are simple. The “closest” function computes the closest mean to a given point x. The “partition” function uses Mathematica’s built-in GatherBy function to partition the points by the closest mean; GatherBy[L, f] partitions its input list L by putting together all points which have the same value under f. The “updatePartition” function creates the new partition based on the centers of the old partition. And finally, the “kMeans” function uses Mathematica’s built-in FixedPoint function to iteratively apply updatePartition to the initial partition until there are no more changes in the output.

Indeed, this is as close as it gets to the “functional” pseudocode we had above. And applying it to some synthetic data (three randomly-sampled Gaussian clusters that are relatively far apart) gives a good clustering in a mere two iterations:

k-means-example

Indeed, we rarely see a large number of iterations, and we leave it as an exercise to the reader to test Lloyd’s algorithm on random noise to see just how bad it can get (remember, all of the code used in this post is available on this blog’s Github page). One will likely see convergence on the order of tens of iterations. On the other hand, there are pathologically complicated sets of points (even in the plane) for which Lloyd’s algorithm takes exponentially long to converge to a fixed point. And even then, the solution is never guaranteed to be optimal. Indeed, having the possibility for terrible run time and a lack of convergence is one of the common features of heuristic algorithms; it is the trade-off we must make to overcome the infeasibility of NP-hard problems.

Our second implementation was in Python, and compared to the Mathematica implementation it looks like the lovechild of MUMPS and C++. Sparing the reader too many unnecessary details, here is the main function which loops the partition updating, a la the imperative pseudocode:

def kMeans(points, k, initialMeans, d=euclideanDistance):
   oldPartition = []
   newPartition = partition(points, k, initialMeans, d)

   while oldPartition != newPartition:
      oldPartition = newPartition
      newMeans = [mean(S) for S in oldPartition]
      newPartition = partition(points, k, newMeans, d)

   return newPartition

We added in the boilerplate functions for euclideanDistance, partition, and mean appropriately, and the reader is welcome to browse the source code for those.

Birth and Death Rates Clustering

To test our algorithm, let’s apply it to a small data set of real-world data. This data will consist of one data point for each country consisting of two features: birth rate and death rate, measured in annual number of births/deaths per 1,000 people in the population. Since the population is constantly changing, it is measured at some time in the middle of the year to act as a reasonable estimate to the median of all population values throughout the year.

The raw data comes directly from the CIA’s World Factbook data estimate for 2012. Formally, we’re collecting the “crude birth rate” and “crude death rate” of each country with known values for both (some minor self-governing principalities had unknown rates). The “crude rate” simply means that the data does not account for anything except pure numbers; there is no compensation for the age distribution and fertility rates. Of course, there are many many issues affecting the birth rate and death rate, but we don’t have the background nor the stamina to investigate their implications here. Indeed, part of the point of studying learning methods is that we want to extract useful information from the data without too much human intervention (in the form of ambient knowledge).

Here is a plot of the data with some interesting values labeled (click to enlarge):

countries-birth-deat-labeled

Specifically, we note that there is a distinct grouping of the data into two clusters (with a slanted line apparently separating the clusters). As a casual aside, it seems that the majority of the countries in the cluster on the right are countries with active conflict.

Applying Lloyd’s algorithm with k=2 to this data results in the following (not quite so good) partition:

countries-birth-death-unstandardized

Note how some of the points which we would expect to be in the “left” cluster are labeled as being in the right. This is unfortunate, but we’ve seen this issue before in our post on k-nearest-neighbors: the different axes are on different scales. That is, death rates just tend to vary more wildly than birth rates, and the two variables have different expected values.

Compensating for this is quite simple: we just need to standardize the data. That is, we need to replace each data point with its deviation from the mean (with respect to each coordinate) using the usual formula:

\displaystyle z = \frac{x - \mu}{\sigma}

where for a random variable X, its (sample) expected value is \mu and its (sample) standard deviation is \sigma. Doing this in Mathematica is quite easy:

Transpose[Map[Standardize, Transpose[L]]]

where L is a list containing our data. Re-running Lloyd’s algorithm on the standardized data gives a much better picture:

countries-birth-death-2cluster

Now the boundary separating one cluster from the other is in line with what our intuition dictates it should be.

Heuristics… The Air Tastes Bitter

We should note at this point that we really haven’t solved the centroid clustering problem yet. There is one glaring omission: the choice of k. This question is central to the problem of finding a good partition; a bad choice can yield bunk insights at best. Below we’ve calculated Lloyd’s algorithm for varying values of k again on the birth-rate data set.

Lloyd's algorithm processed on the birth-rate/death-rate data set with varying values of k between 2 and 7.

Lloyd’s algorithm processed on the birth-rate/death-rate data set with varying values of k between 2 and 7 (click to enlarge).

The problem of finding k has been addressed by many a researcher, and unfortunately the only methods to find a good value for k are heuristic in nature as well. In fact, many believe that to determine the correct value of k is a learning problem in of itself! We will try not to go into too much detail about parameter selection here, but needless to say it is an enormous topic.

And as we’ve already said, even if the correct choice of k is known, there is no guarantee that Lloyd’s algorithm (or any algorithm attempting to solve the centroid clustering problem) will converge to a global optimum solution. In the same fashion as our posts on cryptoanalysis and deck-stacking in Texas Hold ‘Em, the process of finding a minimum can converge to a local minimum.

Here is an example with four clusters, where each frame is a step, and the algorithm progresses from left to right (click to enlarge):

One way to alleviate the issues of local minima is the same here as in our other posts: simply start the algorithm over again from a different randomly chosen starting point. That is, as in our implementations above, our “initial means” are chosen uniformly at random from among the data set points. Alternatively, one may randomly partition the data (without respect to any center; each data point is assigned to one of the k clusters with probability 1/k). We encourage the reader to try both starting conditions as an exercise, and implement the repeated algorithm to return that output which minimizes the objective function (as detailed in the “Partitions and Squared Deviations” section).

And even if the algorithm will converge to a global minimum, it might not be the case that it does so efficiently. As we already mentioned, solving the problem of centroid clustering (even for a fixed k) is NP-hard. And so (assuming \textup{P} \neq \textup{NP}) any algorithm which converges to a global minimum will take exponentially long on some pathological inputs. The interested reader will see this exponentially slow convergence even in the case of k=2 for points in the plane (that is as simple as it gets).

These kinds of reasons make Lloyd’s algorithm and the centroid clustering problem a bit of a poster child of machine learning. In theory it’s difficult to solve exactly, but it has an efficient and widely employed heuristic used in practice which is often good enough. Moreover, since the exact solution is more or less hopeless, much of the focus has shifted to finding randomized algorithms which on average give solutions that are within some constant-factor approximation of the true minimum.

A Word on Expectation Maximization

This algorithm shares quite a bit of features with a very famous algorithm called the Expectation-Maximization algorithm. We plan to investigate this after we spend some more time on probability theory on this blog, but the (very rough) idea is that the algorithm operates in two steps. First, a measure of “center” is chosen for each of a number of statistical models based on given data. Then a maximization step occurs which chooses the optimal parameters for those statistical models, in the sense that the probability that the data was generated by statistical models with those parameters is maximized. These statistical models are then used as the “old” statistical models whose centers are computed in the next step.

Continuing the analogy with clustering, one feature of expectation-maximization that makes it nice is it allows the sizes of the “clusters” to have varying sizes, whereas Lloyd’s algorithm tends to make its clusters have equal size (as we saw with varying values of k in our birth-rates example above).

And so the ideas involved in this post are readily generalizable, and the applications extend to a variety of fields like image reconstruction, natural language processing, and computer vision. The reader who is interested in the full mathematical details can see this tutorial.

Until next time!