Basics of K-Means and DBSCAN Clustering Models for Predictive Analytics
Unsupervised learning has many challenges for predictive analytics — including not knowing what to expect when you run an algorithm. Each algorithm will produce different results; you’ll never be certain whether one result is better than the other — or even whether the result is of any value.
When you know what the outcomes should be, you can tweak the algorithms to produce the desired outcomes. In real-world datasets, you won’t have this luxury. You’ll have to depend on some prior knowledge of the data — or intuition — to decide which initialization parameters and algorithms to use as you create your model.
In real unsupervised learning tasks, however, this prior knowledge is unavailable and the desired result is difficult to find. Choosing the right number of clusters is the key problem. If you happen to stumble upon the right number of clusters, your data will yield insights which you can make highly accurate predictions. On the flip side, guessing the wrong number of clusters may yield subpar results.
K-means algorithm is a good choice for datasets that have a small number of clusters with proportional sizes and linearly separable data — and you can scale it up to use the algorithm on very large datasets.
Think of linearly separable data as a bunch of points in a graph that can be separated using a straight line. If the data is not linearly separable, then more advanced versions of K-means will have to be employed — which will become more expensive computationally and may not be suitable for very large datasets. In its standard implementation, the complexity to compute the cluster centers and distances is low.
K-means is widely employed to solve big-data problems because it’s simple to use, effective, and highly scalable. No wonder most commercial vendors use the K-means algorithm as a key component of their predictive analytics packages.
The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) implementation in scikit-learn does not require any user-defined initialization parameters to create an instance. You can override the default parameters during initialization if you want. Unfortunately, if you’re using the default parameters, the algorithm can’t provide a close match to the desired outcome.
DBSCAN is better suited for datasets that have disproportional cluster sizes, and whose data can be separated in a non-linear fashion. Like K-means, DBSCAN is scalable, but using it on very large datasets requires more memory and computing power.