Tree-based space partition and merging ensemble learning framework for imbalanced problems.

Dealing with imbalanced problems is a significant challenge in machine learning, especially when the data set exhibits an irregular distribution. To this end, this paper proposes a tree-based space partition and merging ensemble learning framework known as the space partition tree (SPT), to partition the data space into two sub-spaces recursively. The partition hyperplane partitions the current space according to the maximum scatter direction of the majority set in the current space. When the partitioned sub-space satisfies the termination conditions, the sub-space is regarded as a decision space, and the decision region of the minority and majority classes is learned in this decision space. By merging the decision regions in all decision spaces, the SPT provides the entire decision region for the original problem. Thereby, the original complex problem can be divided into smaller problems with a relatively balanced and regular distribution. Moreover, the designed partition strategy offers advantages for the recognition of minority samples in the decision space. Finally, the space partition and merging exhibit superior geometric intuition and property of diversity. By introducing the biased penalties Support Vector Machine (BPSVM) into the SPT, the SPT-BPSVM demonstrates satisfactory performance and validates the effectiveness of the SPT in the experiment.