A Comparative Look at the Resilience of Discriminative and Generative Classifiers to Missing Data in Longitudinal Datasets

Longitudinal datasets often suffer from the missing data problem caused by irregular sampling rates and drop-outs, which leads to sub-optimal classification performances. Given the breakthrough of deep generative models in data generation, this paper proposes a conditional Generative Adversarial Network (GAN) based longitudinal classifier (LoGAN) to address this problem in longitudinal datasets. LoGAN is evaluated against commonly used and state-of-art discriminative and generative classifiers. Comparative performance is presented showing the sensitivity of classifiers to data missingness, in both balanced and imbalanced datasets. Results show that the GAN-based models perform on par with other deep learning based models and perform comparatively better on a balanced dataset, while non-deep learning-based discriminative models, in particular the ensemble models, performed better when data was imbalanced. Specifically, F1 scores for LoGAN models were >= 80% for up to 20% of missing data rates in the temporal component of the dataset and >= 60% for missing rates from 40-100%. Non-deep generative models showed low performance with the introduction of missing data rates.