During the knowledge discovery in database process some problems may be found, e.g.
some instance of one attribute may be missing. Such issue can even cause harmful effects to the
final results of the process, since directly affects the data quality of a database which some
machine learning algorithm may be applied to. In the literature are some proposals to solve such
harm; among them is the data imputation process that estimates a plausible value to fill in the
missing one. Inside the area of missing value treatment, some researches were analyzed and
observations were raised such as, a few utilization of synthetic datasets that simulates the main
mechanisms of missingness and a tendency to use bioinspired algorithm to treat the missing
values. From this scenario, the present dissertation analyses an imputation method based on
particle swarm optimization, an underexplored one, and applies it to the treatment of synthetics
datasets generated considering the main mechanisms of missingness, MAR, MCAR and NMAR.
The results obtained when comparing the algorithm against different configurations of itself and
another two treatments known in the area (KNNImpute and SVMImpute) are promising for its
use as missing value treatment whereas the bioinspired method reached the bests values for the
major of the experiments.
|
