Matrix sketching for supervised classification with imbalanced classes

Abstract The presence of imbalanced classes is more and common in practical applications it known to heavily compromise the learning process. In this paper we propose a new method aimed at addressing issue binary supervised classification. Re-balancing class sizes has turned out be fruitful strategy overcome problem. Our proposal performs re-balancing through matrix sketching. Matrix sketching recently developed data compression technique that characterized by property preserving most linear information present data. Such guaranteed Johnson-Lindenstrauss’ Lemma (1984) allows embed an n -dimensional space into reduced one without distorting, within $$\epsilon $$ ? -size interval, distances between any pair points. We use as alternative standard strategies are based on random under-sampling majority or over-sampling minority one. assess properties our when combined with discriminant analysis (LDA), classification trees (C4.5) Support Vector Machines (SVM) simulated real Results show can represent sound widely used rebalancing methods.