Dhaka University Journal of Applied Science & Engineering

While data is integral to address real-world problems using machine learning techniques, the availability of complete data is challenging as data goes missing during collection or even afterward. Missing values make data unsuitable for use and demand imputation techniques to resolve the problem. Here, we compare the performance of four existing missing data imputation techniques KNN, MICE, GAIN, and HexaGAN by applying them on KU-HAR dataset and two datasets of the collection, Nurse Care Activity Recognition Dataset. Our investigation suggests that HexaGAN learns the original data distribution better and demands further considerations, as the RMSE between the imputed data and the real data is lower. To investigate the role of activation function, we replace the underlying ReLU activation functions of the neural networks in HexaGAN architecture with the Swish activation function. Experimental results show that the modified version of HexaGAN possesses the potential to outperform the original one when applied to the same activity recognition datasets.

1. ref1