Modified Fuzzy Dempster-Shafer Theory for Decision Fusion
Somnuek Surathong, Chakkraphop Maisen, Pratch Piyawongwisal · 2021
Decision fusion is a process of combining evidence from multiple classifiers to achieve higher prediction accuracy. Dempster's rule of combination in Dempster-Shafer theory (DST) is one of the most widely-used methods in decision fusion. From previous works, we have developed fuzzy Dempster-Shafer theory (FDST), which incorporates fuzzy set theory into DST by using fuzzy arithmetic to manipulate the uncertainty of data. However, this method still has the following downsides: 1) poor performance in the presence of high conflict of evidence and 2) high time complexity of the fuzzy arithmetic. To address these issues, a new method called modified fuzzy Dempster-Shafer theory (MFDST) is proposed. In this method, the weights of evidence are computed using the distance and entropy of classification outputs. These weights are used in the formula for basic probability assignment (BPA). Additionally, an unnecessary operation in the fuzzy arithmetic is removed, reducing the time complexity. In the experiment, two-classifier fusions based on DST, FDST and MFDST were evaluated on four UCI data sets, using the combinations of the following classifiers: linear discriminant analysis, k-nearest neighbors, naive Bayes, and multilayer perceptron. When comparing MFDST to FDST and DST, we found that MFDST is faster than FDST in all cases. In some cases, MFDST achieves higher accuracy than FDST as well.