Determine the Weight Between Attributes Based on Shapley Value
Lingge Zhou, Bingyi Kang, Jianfeng Zhang · 2021
In the classifier design based on Dempster-Shafer evidence theory (DST), in order to improve the accuracy of classification, a new method to determine evidence fusion among attributes based on the generation of basic probability assignment (BPA) is proposed. The degree of contribution of evidence between different attributes to the fusion results has different weights(reliability), which is calculated based on the Shapley value is proposed in this work. The training data are used to find the best weight for each attribute of the data. Meanwhile, the inter-attribute evidence is discounting according to the best weight and then results are combined using Dempster's combination rule (DCR). Experiments provide the results of classification of Iris data set show that the overall recognition rate is 98.33%. Compared with the previous work, encouraging results are obtained in terms of classification accuracy, which without a large amount of training data and indicated the effectiveness of the proposed method.