cfe8504bda37b575c70ee1a8276f3486-Supplemental.pdf

Namely, we treat the missing indices like a part of imputation targets. We illustrate the extended trainingprocedureinFigure5. E.1 DetailsofimplementationofCSDI We describe the details of architectures and hyperparameters for the conditional diffusion model described in Section 5. First, we provide the whole architecture ofCSDI in Figure 6. As for Transformer layers, we used 1-layer TransformerEncoder implemented in PyTorch [39], which iscomposed ofamulti-head attention layer,fully-connected layers andlayer normalization. For the air quality dataset, following [2], we used the 3rd, 6th, 9th and 12th months as test data.