Missing value imputation in Food Composition Data with Denoising Autoencoders

Abstract

Missing data is a common problem in a wide range of fields that can arise as a result of different reasons: lack of analysis, mishandling samples, measurement error, etc. The area of nutrition and food composition is no exception to the problem of missing values. Missing data in food composition databases (FCDB) significantly limits their usage. Commonly this problem is resolved by calculating mean or median from available data in the same FCDB or borrowing values from other FCDBs, however, this method produces notable errors. This paper focuses on missing value imputation using autoencoders, a deep learning algorithm that has the ability to approximate values by learning a higher-level representation of its input. The data used was from the FCDBs collected by the USDA FoodData Central. We compared the autoencoder imputation method with the commonly used approaches fill-in-with-mean and fill-in-with-median, and the results show that the autoencoder method for imputation provides superior results.