基于构造性神经网络与全局密度信息的不平衡数据欠采样方法

Imbalanced Undersampling Based on Constructive Neural Network and Global Density Information

多数类欠采样是当前数据层面解决不平衡数据学习的主流技术之一,近年来,研究者们提出了一系列的欠采样方法,但大多都将重点放在如何选择代表性的样本,从而降低信息损失。然而,如何在欠采样过程中保持多数类内部的结构信息,仍然是欠采样面临的主要挑战。针对该挑战,提出了一种基于构造性神经网络和全局分布密度的不平衡数据集欠采样方法。该方法首先基于构造性神经网络,设计了一种多数类局部模式的学习方法;然后基于多数类局部模式,设计了两种具有结构保持特性的样本选择策略;最后针对局部模式学习的随机性可能导致的采样结果非优的问题,进一步引入了bagging集成策略,提升了方法的性能。在59个数据集上与13种对比方法进行了对比实验,验证了所提方法在G-mean,AUC和F1-score这3个常用指标上的有效性。

Undersampling is one of the mainstream data-level technologies to deal with imbalanced data.In recent years,researchers have proposed numerous undersampling methods,but most of them focus on how to select representative majority class samples to avoid the loss of informative data.However,how to maintain the structures of the original majority class in the process of undersampling is still an open challenge.To this end,an undersampling method for imbalanced data classification is proposed based on constructive neural network and data density.Firstly,it detects the majority local patterns with a simplified constructive process.Then,two sample selection strategies are designed to maintain the structure of the selected groups according to the original majority distribution information.Finally,to solve the problem that the randomness of local pattern learning may lead to non-optimal sampling results,the bagging technique is introduced to further improve the learning performance.Comparative experiments with 13 comparison methodson 59 datasets verify the effectiveness of the proposed method in terms of three metrics G-mean,AUC and F1-score.