Unsupervised Representation Learning

To compare the performance of the representation learning methods, we generated embeddings of all clinical concepts (nodes) in the SemMedDB using DeepWalk, TransE, Weighted DeepWalk, Weighted TransE, and ESP. We used the available implementation of DeepWalk [33] and TransE [34]. We experimented with multiple hyperparameter sets and selected the ones that yielded the best loss value during representation learning. In Weighted TransE and TransE, we set the α value to .001, batch size to 256 triplets, epochs to 100, and the number of corrupted triplets for each positive triplet to 1. The embedding dimension was 100 in both these models. For the DeepWalk and Weighted DeepWalk models, we set the walk length to 500, the number of walks to 20, window size to 4, α value to .025, and an embedding size of 256. For ESP, we used the embeddings provided by the authors [35], which had a dimension of 8000.