Hello Alfredo,
I work in Antonio’s lab and am using the HI-VAE for a project in Pablo’s graduate deep learning course. I am trying this out with heterogenous medical data (35 features) with true missing values to predict a Multilabel of has disease (0 or 1) and type of care unit (ordinal 0,1,2,). I have about 80% of the dataset with labels and 20% without labels (I am holding onto these labels and using them to test the output of the HI-VAE), while both still contain true missing value. I have the model working with randomly generated masks that never coincide with the true missing values and then the csv of true missing values. Do you have any tips or suggestions for improving accuracy? Also I am unsure if the final result I am getting is assuming the least amount missing as it should for the final prediction.
Hello Alfredo,
I work in Antonio’s lab and am using the HI-VAE for a project in Pablo’s graduate deep learning course. I am trying this out with heterogenous medical data (35 features) with true missing values to predict a Multilabel of has disease (0 or 1) and type of care unit (ordinal 0,1,2,). I have about 80% of the dataset with labels and 20% without labels (I am holding onto these labels and using them to test the output of the HI-VAE), while both still contain true missing value. I have the model working with randomly generated masks that never coincide with the true missing values and then the csv of true missing values. Do you have any tips or suggestions for improving accuracy? Also I am unsure if the final result I am getting is assuming the least amount missing as it should for the final prediction.