Cohesion Metrics

Since the Stevens–Myers–Constantine taxonomy of 1974, cohesion has been measured through structural proxies: method–attribute access overlap, connected components of the call graph, data-slice token sharing, and normalized densities. This paper surveys the 24 major cohesion metrics in the literature and shows that every single one measures a structural coupling relation, not cohesion as defined by a modularization principle. The underlying input to every metric is a coupling graph — method pairs that share fields, elements that share data tokens, execution traces — and what distinguishes “cohesion” metrics from classical coupling metrics (CBO, RFC, MPC) is not the data but the scope: coupling metrics count edges that cross module boundaries, while “cohesion” metrics count edges that stay inside. Both families examine the same graph; they differ only in which partition of the edge set they report. The paper catalogs each metric, identifies the coupling relation it extracts and the transformation it applies, and demonstrates that none captures the property it claims to measure. The finding is definitional, not pragmatic: structural scores remain useful as screening heuristics and change predictors, but they cannot be interpreted as measures of principled cohesion.