Estimating Optimal Values for Intentional-Value-Substitution Learning

Takuya Fukushima, Taku Hasegawa, Tomoharu Nakashima

Abstract: Intentional-Value-Substitution (IVS) learning was shown to be effective for missing data. This paper gives further insight into the optimal value for the substitution during the learning of a regression model. Function fitting is focused on as the task of the machine learning model. Theoretical analysis on the optimal substitution value for IVS learning is presented before a series of experiments with neural networks are conducted in order to confirm the validity of the theoretical analysis. This paper also proposes a method for estimating the optimal substitution values without using the information of the target function. Another series of computational experiments are conducted to evaluate the accuracy performance of the estimation method.