Author: Kevin Feasel

Matthew Mayo shares a few tips:

The k-means algorithm is a cornerstone of unsupervised machine learning, known for its simplicity and trusted for its efficiency in partitioning data into a predetermined number of clusters. Its straightforward approach — assigning data points to the nearest centroid and then updating the centroid based on the mean of the assigned points — makes it one of the first algorithms most data scientists learn. It is a workhorse, capable of providing quick and valuable insights into the underlying structure of a dataset.

This simplicity comes with a set of limitations, however. Standard k-means often struggles when faced with the complexities of real-world data. Its performance can be sensitive to the initial placement of centroids, it requires the number of clusters to be specified in advance, and it fundamentally assumes that clusters are spherical and evenly sized. These assumptions rarely hold true in the wild, leading to suboptimal or even misleading results.

Read on for a few ways to relax some of the constraints in k-means clustering.

Ivan Palomares Carrascosa continues a series on decision trees:

But what are the underlying mechanisms that make decision trees so well-suited for various predictive tasks? And what criteria are internally used to construct them? Specifically, how are nodes recursively split as the tree-shaped structure is formed? This article takes a closer look at the inner workings of decision trees, focusing on how branches are created through deliberate, data-driven splitting (spoiler: it certainly doesn’t happen at random).

One of the main principles of CART is around finding efficient splits for trees, and this digs into some of those details.