Performance of ACO-based Decision Tree Algorithm with Imbalanced Class Data Sets - A Heuristic Approach

The prediction of minority class can be like finding a needle in a haystack. Bio-inspired classifier such as Ant Colony Optimization (ACO) decision tree experienced ineffective decision boundaries since its entropy-based heuristic is affected by the strong presence of the dominant class. Consequently, the developed trees were dominated by the likelihood of the majority class where the rare class is under-represented. The proposed algorithm with class skew-insensitive heuristic namely the Hellinger-Ant-Tree-Miner (HATM) was compared to the Ant-Tree-Miner (ATM), via a simulation study and application to 15 imbalanced data. Simulation results revealed the advantage of HATM over the ATM under skewed class distributions as the number of covariates and sample sizes increase. Experiments with real data indicate a potential improvement of the ATM measured by balanced accuracy (BACC), F-Measure and minority class prediction (MCP). The Friedman tests justify that HATM performed better than ATM while being competitive with other well-known tree-based classifiers.