深度置信网络的代价敏感多粒度三支决策模型研究

最优粒度选择是自编码网络构造多粒度特征的关键环节。针对自编码网络粒度选择方法不合理导致特征提取效果差以及错误分类成本和测试成本高的问题,提出一种基于小批量梯度下降(mini-batch gradient descent, MBGD)的粒度层选取策略。该方法通过改变粒度选择方式重新构建多粒度空间,设计一个新的基于深度置信网络(deep belief network, DBN)的代价敏感多粒度三支决策模型。更优的粒度选择方法提升网络的特征提取能力,促使多粒度空间的构造朝着最快到达最细粒度空间的方向发展,降低图像重构误差以达到更小的错误分类代价和测试代价。实验结果表明,提供合理的粒度选取策略提高了代价敏感多粒度三支决策模型的决策准确性,并在给定代价情况下更快地获得总代价最小的最优粒层。

Research on cost-sensitive multi-granularity three-way decision model for deep belief networks

Optimal granularity selection is a critical link in constructing multi-granularity features in autoencoder networks. Aiming at the problems of unreasonable granularity selection method in autoencoder networks, which led to unsatisfactory feature extraction effect, high misclassification cost, and high testing cost, this paper proposed a granularity layer selection strategy based on mini-batch gradient descent(MBGD). This approach reconstructed the multi-granularity space by changing the granularity selection method and designing a new cost-sensitive multi-granularity three-way decision model based on a deep belief network(DBN). The superior granularity selection method enhanced the feature extraction ability of the network, promoted the construction of multi-granularity space to develop towards the fastest reaching the most fine-grained area, and reduced the image reconstruction error to achieve more minor misclassification cost and test cost. The experimental results show that providing a reasonable granularity selection strategy improves the decision accuracy of the cost-sensitive multi-granularity three-way decision model, and can obtain the optimal granular layer with the smallest total cost faster under the given cost.