A Method for Improving the Accuracy of Regression Models Based on Ordinal-Invariant Pattern Clustering

Procedia Computer Science. 2025. Vol. 266. P. 1330–1335.

Alexey Myachin

An agglomerative pattern analysis algorithm is presented that groups objects using ordinal-invariant pattern clustering, with the goal of maximizing the coefficient of determination. Key stages are described: constructing initial pattern pattern by grouping objects according to permutations of feature values; computing centroids and variances; and defining a merge criterion based on the maximal increase in R2. Algorithmic details include incremental updates of cumulative statistics, enabling constant-time updates per merge. A complexity analysis shows that the basic implementation is cubic in the number of initial clusters, but filtering and priority queues can reduce the practical running time. Experimental evaluation on a synthetic dataset and on the red wine quality dataset from the UCI Machine Learning Repository compares the proposed method with k-nearest neighbours regression and standard regression methods, demonstrating consistently higher R2 scores.

Keywords: regression models, pattern analysis, ordinal-invariant pattern clustering