We currently have methods to compute feature importance for supervised tasks (e.g. for each target/survival variable). We need to implement feature importance for autoencoders/cross-modality encoders for the encoder/decoder connections. What is the subset of features in input layers that are most informative for reconstructing each feature in the output layers?
We currently have methods to compute feature importance for supervised tasks (e.g. for each target/survival variable). We need to implement feature importance for autoencoders/cross-modality encoders for the encoder/decoder connections. What is the subset of features in input layers that are most informative for reconstructing each feature in the output layers?