This is the source code for paper Cola-GNN: Cross-location Attention based Graph Neural Networks for Long-term ILI Prediction appeared in CIKM2020 (research track)
The raw dataset are in in the data folder. For each dataset, there are two files defined. For example, for the Japan-prefecture dataset, we have two files:
japan.txtincludes the spatiotemporal data. Columns indicate locations (i.e., prefecture) and rows indicate timestamps (i.e., weeks). Each value is the number of patients in a location at a time point. The data are arranged in chronological order.japan-adj.txtcontains a adjacency matrix.
The training data are processed by the DataBasicLoader
class in the src/data.py file. We can set different value for historical window size args.window and horizon/leadtime args.horizon. Setting args.window=20, args.horizon=1/2 means using data from the previous 20 weeks to predict the upcoming/next week. There are some functions in this class:
- _split splits the data into training/validation/test sets.
- _batchify generates data samples. Each sample contains a time series input with length equal to args.window, and a value for the output. For the current code, there are overlaps in the inputs of different samples.
- get_batches generates random mini-batches for training.