Imputing incomplete multi-dimensional data using neural network and clustering similarity comparison

Abstract

This dissertation presented a method to fill in missing data in multi-dimensional data. These data are divided into two categories. The first one is incomplete time series data. The algorithm for imputing the missing time-series data is based on the gradient of the area surrounding the missing data. The missing information which is the gradient of a data falls in one of the following three categories: positive gradient, negative gradient, and zero gradient. When a group of missing data belongs to one of three categories, the missing data are imputed with bootstrapping method. The second type is filling in the incomplete multi-dimensional data in an image. To impute the missing image, the characteristics of missing image are used. If missing data are randomly and fine scattered, an artificial neural network model is used to create an approximated surface to cover those missing data. But if the missing data are clustered in forms of an empty shape, then a similarity pattern searching and filling is performed. The missing data areas are divided into a set of equal size of windows. This windowed area will be compared with every other non-missing data area of the image area to find the most similar area with the missing area. The experimental results concluded that our proposed algorithms are outperformed the other tradition methods in several cases.