Awesome

TIDER

Code for paper titled 'Multivariate Time-series Imputation with Disentangled Temporal Representations'.

Code is written in PyTorch v1.9.0+cu111. Python version is Python 3.6.9.

Simply 'python3 TIDER.py' can conduct the training, validation and testing process.

##Detailed explanations of hyperparameters: • --save_path: address to save the optimal model parameters.

• --datadir: address of time-series data.

• --device: cpu / gpu device.

• --valid: proportion of validation data.

• --drop_rate: data removing rate

• --eta: hyperparameter for L2 regularization

• --n_test: testing set temporal length

• --num_epochs: number of epochs

• --batch_size: batch size

• --dim_size: dimension of feature matrix

• --lag_list: W of bias matrix

• --lambda_ar: weight for bias matrix constraint function

• --lambda_trend: weight for trend matrix constraint function

• --bias_dimension: dimension for bias feature matrix

• --season_num: K

• --seasonality: seasonality of time-series

• --learning_rate: learning rate

• --lambda_trend: control the loss function of trend matrix

For Guangzhou data: It is a 21461144 tensor, we can first transfer it into a 214*8784 matrix, then use the last 500 columns as the processed matrix

For Westminster data: a csdn blog is close to our processed: https://blog.csdn.net/qq_40206371/article/details/128932640

nearly same as TIDER. The differences lies in:

• no hyperparameter 'seasonality'

• lambda_spatial: control the Laplacian regularization term for matrix U

• topk_freq: the number of frequencies selected in multi-periods seasonality matrix.