Discrete Hierarchical Conditional Variational Auto-Encoder and Self Attention Mechanism

The evaluations are modified with the feed back mechanism based on optimal model in Large Scale Group Decision Making (LSGDM) usually, the intelligent decision making cannot be achieved with end-to-end. The application of LSGDM is limited, such as the customer evaluation to sales factors, the most customers would not modify the provided evaluations. A novel method combining Conditional Variational Auto-Encoder (CVAE) and self attention mechanism is developed to conduct the intelligent decision making with end-to-end. After the evaluators provide the evaluations through pair-wise comparison of factors using Lginstic Preference Relation (LPR), the authors first construct the Discrete Hierarchical Conditional Variational Auto-Encoder (DHCVAE) model to generate the matrices with high consistency level according to the transformed matrices from the individual evaluations, and the quality of generated matrix is also improved. Second, the authors develop the method to aggregate the generated matrices to obtain the final consensus matrix via self attention mechanism. In the process of constructed self attention, all matrices would nearly reach consensus. The method deploys the action-state transition and policy gradient optimization to converge the process of constructed self attention quickly. The case study is conducted to verify the advantages of the developed method and the comparisons with other state-of-arts are also conducted. This data set is used in the case study.