Density-Based Spatial Clustering of Applications with Noise (DBSCAN)
DBSCAN is a density-based clustered algorithm similar to mean-shift, but with a couple of notable advantages. Check out another fancy graphic below and let’s get started!
DBSCAN Smiley Face Clustering
- DBSCAN begins with an arbitrary starting data point that has not been visited. The neighborhood of this point is extracted using a distance epsilon ε (All points which are within the ε distance are neighborhood points).
- If there are a sufficient number of points (according to pinpoints) within this neighborhood then the clustering process starts and the current data point becomes the first point in the new cluster. Otherwise, the point will be labeled as noise (later this noisy point might become part of the cluster). In both cases that point is marked as “visited”.
- For this first point in the new cluster, the points within its ε distance neighborhood also become part of the same cluster. This procedure of making all points in the ε neighborhood belong to the same cluster is then repeated for all of the new points that have been just added to the cluster group.
- This process of steps 2 and 3 is repeated until all points in the cluster are determined i.e all points within the ε neighborhood of the cluster have been visited and labeled.
- Once we’re done with the current cluster, a new unvisited point is retrieved and processed, leading to the discovery of a further cluster or noise. This process repeats until all points are marked as visited. Since at the end of this all points have been visited, each point will have been marked as either belonging to a cluster or being noisy.
DBSCAN poses some great advantages over other clustering algorithms. Firstly, it does not require a pe-set number of clusters at all. It also identifies outliers as noises, unlike mean-shift which simply throws them into a cluster even if the data point is very different. Additionally, it can find arbitrarily sized and arbitrarily shaped clusters quite well.
The main drawback of DBSCAN is that it doesn’t perform as well as others when the clusters are of varying density. This is because the setting of the distance threshold ε and pinpoints for identifying the neighborhood points will vary from cluster to cluster when the density varies. This drawback also occurs with very high-dimensional data since again the distance threshold ε becomes challenging to estimate.
Expectation-Maximization (EM) Clustering using Gaussian Mixture Models (GMM)
One of the major drawbacks of K-Means is its naive use of the mean value for the cluster center. We can see why this isn’t the best way of doing things by looking at the image below. On the left-hand side, it looks quite obvious to the human eye that there are two circular clusters with different radius’ centered at the same mean. K-Means can’t handle this because the mean values of the clusters are very close together. K-Means also fails in cases where the clusters are not circular, again as a result of using the mean as cluster center.
Two failure cases for K-Means
Gaussian Mixture Models (GMMs) give us more flexibility than K-Means. With GMMs we assume that the data points are Gaussian distributed; this is a less restrictive assumption than saying they are circular by using the mean. That way, we have two parameters to describe the shape of the clusters: the mean and the standard deviation! Taking an example in two dimensions, this means that the clusters can take any kind of elliptical shape (since we have a standard deviation in both the x and y directions). Thus, each Gaussian distribution is assigned to a single cluster.
To find the parameters of the Gaussian for each cluster (e.g the mean and standard deviation), we will use an optimization algorithm called Expectation-Maximization (EM). Take a look at the graphic below as an illustration of the Gaussians being fitted to the clusters. Then we can proceed with the process of Expectation-Maximization clustering using GMMs.
EM Clustering using GMMs
- We begin by selecting the number of clusters (like K-Means does) and randomly initializing the Gaussian distribution parameters for each cluster. One can try to provide a good estimate for the initial parameters by taking a quick look at the data too. Though note, as can be seen in the graphic above, this isn’t 100% necessary as the Gaussians start as very poor but are quickly optimized.
- Given these Gaussian distributions for each cluster, compute the probability that each data point belongs to a particular cluster. The closer a point is to the Gaussian’s center, the more likely it belongs to that cluster. This should make intuitive sense since with a Gaussian distribution we are assuming that most of the data lie closer to the center of the cluster.
- Based on these probabilities, we compute a new set of parameters for the Gaussian distributions such that we maximize the probabilities of data points within the clusters. We compute these new parameters using a weighted sum of the data point positions, where the weights are the probabilities of the data point belonging to that particular cluster. To explain this visually we can take a look at the graphic above, in particular the yellow cluster as an example. The distribution starts randomly on the first iteration, but we can see that most of the yellow points are to the right of that distribution. When we compute a sum weighted by the probabilities, even though there are some points near the center of investment that converge, most of them are on the right. Thus naturally the distribution’s mean is shifted closer to that set of points. We can also see that most of the points are “top-right to bottom-left”. Therefore the standard deviation changes to create an ellipse that is more fitted to these points, to maximize the sum weighted by the probabilities.
- Steps 2 and 3 are repeated iteratively until convergence, where the distributions don’t change much from iteration to iteration.
There are 2 key advantages to using GMMs. Firstly GMMs are a lot more flexible in terms of cluster covariance than K-Means; due to the standard deviation parameter, the clusters can take on any ellipse shape, rather than being restricted to circles. K-Means is a special case of GMM in which each cluster’s covariance along all dimensions approaches 0. Secondly, since GMMs use probabilities, they can have multiple clusters per data point. So if a data point is in the middle of two overlapping clusters, we can simply define its class by saying it belongs X-percent to class 1 and Y-percent to class 2. I.e GMMs support mixed membership.
Agglomerative Hierarchical Clustering
Hierarchical clustering algorithms fall into 2 categories: top-down or bottom-up. Bottom-up algorithms treat each data point as a single cluster at the outset and then successively merge (or agglomerate) pairs of clusters until all clusters have been merged into a single cluster that contains all data points. Bottom-up hierarchical clustering is therefore called hierarchical agglomerative clustering or HAC. This hierarchy of clusters is represented as a tree (or dendrogram). The root of the tree is the unique cluster that gathers all the samples, the leaves being the clusters with only one sample. Check out the graphic below for an illustration before moving on to the algorithm steps
Agglomerative Hierarchical Clustering
- We begin by treating each data point as a single cluster i.e if there are X data points in our dataset then we have X clusters. We then select a distance metric that measures the distance between two clusters. As an example, we will use average linkage which defines the distance between two clusters to be the average distance between data points in the first cluster and data points in the second cluster.
- On each iteration, we combine two clusters into one. The two clusters to be combined are selected as those with the smallest average linkage. I.e according to our selected distance metric, these two clusters have the smallest distance between each other and therefore are the most similar and should be combined.
- Step 2 is repeated until we reach the root of the tree i.e we only have one cluster which contains all data points. In this way, we can select how many clusters we want in the end, simply by choosing when to stop combining the clusters i.e when we stop building the tree!
Hierarchical clustering does not require us to specify the number of clusters and we can even select which number of clusters looks best since we are building a tree. Additionally, the algorithm is not sensitive to the choice of distance metric; all of them tend to work equally well whereas, with other clustering algorithms, the choice of distance metric is critical. A particularly good use case of hierarchical clustering methods is when the underlying data has a hierarchical structure and you want to recover the hierarchy; other clustering algorithms can’t do this. These advantages of hierarchical clustering come at the cost of lower efficiency, as it has a time complexity of O(n³), unlike the linear complexity of K-Means and GMM