Clustering Large Data Sets with Mixed Numeric and Categorical Values

The paper defines a hard-clustering objective whose numerical portion is minimized by cluster means and categorical portion by modal values. Allocation and prototype updates repeat until no records move, inheriting k-means' efficiency and local-optimum dependence.

Durable contribution #

The important mechanism is not merely “k-means for categorical data.” It is a composite dissimilarity:

$$d(x,q)=\sum_{j\in\mathcal N}(x_j-q_j)^2 +\gamma\sum_{j\in\mathcal C}\mathbf1(x_j\ne q_j).$$

The weight $\gamma$ decides how a categorical mismatch trades against numerical distance. The paper gives standard-deviation-based guidance and explicitly says more work is needed; that heuristic should not be promoted to a universal setting rule.

Limits #

Every unequal category pair has the same cost, numerical scaling remains decisive, initialization affects the local optimum, and the original runtime measurements are hardware-specific historical results. Cluster descriptions induced afterward by a decision tree describe the computed partition; they do not validate it as a natural population structure.

Visual audit #

All fourteen pages were rendered and inspected. The weight-effect sketches, pseudocode, convergence curve, synthetic partitions, and runtime tables were reviewed. None is retained: the composite objective plus a worked preprocessing explanation is clearer at article width, and historical runtime plots are not current evidence.