anticlust

anticlust partitions a pool of elements into anticlusters by maximizing heterogeneity within subsets to produce partitions that are mutually similar for applications such as balanced assignments and cross-validation.

Key Features:

• Reversal of clustering objectives: Reverses traditional clustering objectives such as intra-cluster variance (as used in k-means) and the sum of pairwise distances within clusters to maximize within-subset heterogeneity.
• Dual anticlustering criteria: Implements two primary anticlustering criteria by reversing the methodologies of k-means clustering and cluster editing.
• Implementation: Provided as an open-source extension to the R programming language.
• Performance: Simulation studies reported superior performance relative to random assignment and matching in producing balanced partitions.

Scientific Applications:

• Educational assignment: Assigns students to parallel courses to balance groups in terms of skills or knowledge levels.
• Experimental psychology: Assembles equivalent stimulus sets (e.g., from norming data) across experimental conditions to support validity and reliability.
• Test construction: Splits achievement tests into parts with equal item difficulty and discrimination.
• Cross-validation data partitioning: Divides large datasets into subsets for cross-validation while maintaining similar statistical properties across subsets.

Methodology:

Computationally reverses k-means and cluster-editing objectives by maximizing intra-cluster variance or the sum of pairwise distances within clusters, implementing two anticlustering criteria derived from those methods.