Application of similarity optimization deep forest in the classification of unbalanced datasets
Hongwei Chen, Fengzhou Zhang, Zihong Peng, Han Wang · Research Square · 2023
Abstract In fields such as computer vision, information security, and healthcare, data classification faces the challenge of uneven data distribution. To address this problem, the academic community has proposed the Deep Forest algorithm, which combines random forest and extremely random forest techniques. However, in random forests, there are differences in the classification performance of decision trees, and it is not possible to control the diversity of decision trees, which may affect the algorithm’s performance. This paper proposes a method called LCHCDF (Low-Coupling and High-Cohesion Deep Forest) classifier to address these issues, which uses AUC evaluation and inner product comparison strategies to handle them. Experimental validation on UCI datasets shows that LCHCDF has good classification accuracy and scalability and performs well in large-scale imbalanced network attack datasets. In addition, we also propose a distributed improvement algorithm that divides the random forest into sub-forests for parallel computation. The experimental results demonstrate that the LCHCDF algorithm performs well in terms of scalability and classification accuracy.