Page 33 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)
P. 33

Bao Ngoc Vi, Cao Truong Tran, Chi Cong Nguyen                                    17



                                                                                                     (1)



                     The generator   take  ,      and a  -dimension noise variable    as input and outputs
                     a vector of imputations   :
                                                                                                     (2)
                     where     is sampled from a known distribution such as normal distribution, and    is
                     element-wise multiplication.
                        outputs a value for every component as long as its value is observed. Therefore the
                     final completed data vector is obtained by replacing missing value in    with the cor-
                     responding value of   , that is:
                                                                                                     (3)



                     3.2   Discriminator

                     The discriminator attempts to distinguish which components of completed vector
                     are real (observed) or fake (imputed). That is, it try to predicting the mask vector   .
                     In [21], authors showed that if the discriminator    is not provided "enough" infor-
                     mation about     ,    could reproduce several populations that would all be optimal
                     with respect to   . Thus it is necessary to introduce a hint mechanism. A hint mecha-
                     nism is a random variable    and contains some information about M to guarantee that
                     the generator learns the desired distribution.     is defined as below:
                                                                                                     (4)

                     where                              is  defined  by first sampling k from       uni-
                     formly at random and then setting:

                                                                                                     (5)


                        and    are fed to    to predict which components are real or imputed.    is denoted
                     as output of   , that is:

                                                                                                     (6)
                     The discriminator is updated after each evolutionary step to further to distinguish which
                     components of completed vector     are real (observed) or fake (imputed), that is:
                                                                                                     (7)




                     3.3   Evolutionary algorithm



                     Variation Given an individual     in the population, the variation operators is utilized
                     to produce its offspring                 . Specifically, several copies of each parent





                     ISBN: 978-604-80-8083-9                                                  CITA 2023
   28   29   30   31   32   33   34   35   36   37   38