Fig. 5From: Interpretable instance disease prediction based on causal feature selection and effect analysisThe instance is input to the selector network, which outputs the selection probability vector. The selection vector is then sampled based on these probabilities. Then, the prediction network receives the selected features and makes predictions, and the baseline network gives the entire feature vector and makes predictions. Each of these networks is back-propagated training using real labels. Then subtract the loss of the baseline network from the loss of the prediction network, which is used to update the selector network. CPN counterfactual prediction network, CSN counterfactual selection network, FPN fact prediction networkBack to article page