Just like K-Nearest Neighbours and regression models, unsupervised clustering algorithms such as K-Means and Hierarchical Clustering can fall into similar traps of erasure when exposed to lopsided historical data.
K-Means Clustering and Hierarchical Clustering
K-Means works by partitioning data points into k distinct clusters. It randomly initialises k points, called centroids, calculates the distance (usually Euclidean) from every data point to these centroids, and then updates the centroids iteratively by minimising the Within-Cluster Sum of Squares (WCSS).
K-Means is heavily influenced by the distribution of the data. Because the internet is flooded with millions of Western data points, those points can form massive, tightly packed, high-density clouds in vector space.
When the algorithm calculates the average position of these points—the centroid—the sheer volume of Western data can anchor the centroid firmly in “Eurocentric space.”
When the model encounters a unique data profile, that data point may sit in a sparse, low-density region of the vector space. Because K-Means forces every data point into one of the k clusters in order to minimise overall within-cluster variance, a unique point may be assigned to the nearest large Western cluster.
The algorithm’s centroid effectively “swallows” this unique reality.
Hierarchical Clustering
Hierarchical Clustering, specifically Agglomerative Clustering (which is bottom-up), builds a tree-like diagram called a dendrogram by continuously merging the two closest data points or clusters according to a specific linkage criterion, such as Single, Complete, or Ward’s linkage.
If the algorithm uses Single Linkage (Minimum Distance), it can suffer from chaining. A chain of intermediate, poorly documented, or generalised data points can bridge the gap between unique data and a massive Western data cloud, causing the unique narrative to be absorbed into a larger cluster.
If it uses Complete Linkage (Maximum Distance) or Ward’s Method (Variance Minimisation), the resulting clustering can favour compact, well-represented groups. Because unique data are underreported, they may lack the dense, step-by-step structural data points needed to form their own robust branches in the hierarchy.
When the algorithm builds the hierarchy, these unique data points may either appear as isolated “anomalies” at the top of the tree—where engineers may subsequently treat them as noise—or be forced into an ill-fitting sub-branch of a Western cluster, potentially erasing the systemic differences that made them unique in the first place.
Conclusion
While clustering is unsupervised and relies heavily on geometric relationships such as distance and density, it is not inherently biased towards any particular culture.
However, in a lopsided data ecosystem, these algorithms can effectively treat the dominant Eurocentric data as the standard geography of the vector space, mechanically marginalising or absorbing unique non-Western data points.
To force clustering algorithms to better represent alternative narratives, engineers may need to alter the distance metrics—moving beyond simple spatial distance towards metrics or representations that give greater weight to alternative structural and systemic variables.
Day 26 / 30 of the #AIRewardmaxxing Series.
Tomorrow in Part 27, we are diving into Time Series analysis and Association Rule learning to see how temporal and transactional models compound algorithmic bias.
What's your take? Have you seen Al algorithms categorize unique cultural or local edge cases as "statistical noise" or anomalies? Let's discuss below!