Building Autocorrelation-Aware Rerpresentations for Fine-Scale Spatiotemporal Prediction

Y. Lin, Y. -Y. Chiang, M. Franklin, S. P. Eckel and J. L. Ambite, "Building Autocorrelation-Aware Representations for Fine-Scale Spatiotemporal Prediction," 2020 IEEE International Conference on Data Mining (ICDM), 2020, pp. 352-361, doi: 10.1109/ICDM50108.2020.00044.

Requirements

• See torch-env.yml under env folder.

Datasets

Sample data is available at: sample_data/los_angeles_500m_2020_02.npz contains feature values

• dynamic_feature_names (list): list of dynamic features, such as 'temperature', 'dew_point', 'humidity', etc.
• static_feature_names (list): list of static features, such as 'landuse_a_allotments', 'landuse_a_military', etc.
• mapping_map (ndarray): (Height, Width) array of the pixels

See other datasets: https://drive.google.com/drive/folders/1-NO-h--nC2QtJ8lpaC0xX_qGbLUBh9sU?usp=sharing

Running the code

You can find a more detail in "*python train.py -h"

Some commonly-used paramters:

• [--data_path DATA_PATH]
• [--result_dir RESULT_DIR]
• [--model_dir MODEL_DIR]
• [--model_name MODEL_NAME]
• [--model_types MODEL_TYPES]
• [--device DEVICE]
• [--num_epochs NUM_EPOCHS]
• [--batch_size BATCH_SIZE]
• [--lr LR]
• [--weight_decay WEIGHT_DECAY]
• [--patience PATIENCE]
• [--seq_len SEQ_LEN]
• [--en_features EN_FEATURES]
• [--de_features DE_FEATURES]
• [--kernel_sizes KERNEL_SIZES]
• [--h_channels H_CHANNELS]
• [--fc_h_features FC_H_FEATURES]
• [--sp_neighbor SP_NEIGHBOR]
• [--tp_neighbor TP_NEIGHBOR]
• [--alpha ALPHA]
• [--beta BETA]
• [--gamma GAMMA]
• [--eta ETA]
• [--use_tb]
• [--tb_path TB_PATH]
• [--verbose]

Data format

The input includes the contextual data and the available measurements in a grid structure (tensors). The contextual data is a single array of shape H x W x P, where P is the number of input features, H and W are the height and width of the grid. Each input represents the input signal at time i, and T' is the number of previous hours (from t - T' + 1 to t).

Code

Code structure

• The PyTorch implementation of DeepLatte architecture is located in the models folder
• The folder models contains autoencoder, convlstm, and linear layers with other utilites
• DeepLatte model contains two high-level python scripts at train.py and test.py

Reference

• ConvLSTM has been reproduced in DeepLatte following by ConvLSTM.pytorch (Ref: https://github.com/spacejake/convLSTM.pytorch/blob/master/convlstm.py).