Learning using rebalanced statistical invariants for imbalanced classification
Zhen-Qi Liu, Yuan‐Hai Shao · Procedia Computer Science · 2022
Traditional classification algorithms can be limited in their performance on highly imbalanced data sets, resulting in a drop in the classification performance on the minority class. The recent proposed learning using statistical invariants (LUSI) provides a new learning paradigm to the problem of classification. In order to resolve the issues of class imbalance, inspired by the idea of the prior data information, we compute the appropriate predicates and construct the corresponding invariants in LUSI. Additionally, we incorporated different ”rebalance” heuristics in LUSI modeling, which give sufficient consideration to the importance of the probability label information in classification and combine with the invariants to construct rebalanced statistical invariants in rebalancing environment, that effectively solves the imbalanced block of V-Matrix in LUSI. Experimental results signify the effectiveness of appropriate predicate for imbalanced data, in addition, one can see that the proposed rLUSI demonstrated superior performance in comparison to LUSI in dealing with imbalance problem.