In this work, we introduce subarchetypal analysis, a new semi-supervised framework for Archetype Analysis (AA) and Archetypoid Analysis (ADA) based on subsets of observations from which archetypal profiles are generated. To the best of our knowledge, this is the first subset-constrained framework for archetypal analysis. Classical AA and ADA are recovered as special cases, and the location result for archetypes is extended to the proposed framework. The generating subsets may be defined a priori by the user to incorporate external knowledge, or determined from intrinsic properties of the data or the underlying problem structure. As an illustration of the latter, we consider efficient subarchetypes within Data Envelopment Analysis, where the subsets are determined by the geometry of the efficient frontier. We develop and implement algorithms for subarchetypal analysis in both multivariate and functional settings, available in an R package. The proposed methods are compared with constrained clustering, group-wise archetypal analysis, and classical unconstrained AA and ADA, and their practical usefulness is demonstrated through applications to business, sports analytics, and human development data.
翻译:暂无翻译