Social-Aware Time Series Imputation

This project implements the social-aware time series imputation method proposed in [1], which is an imputation algorithm for time series data in the social network.

Testing

This project is implemented in Python 3.6

Dependency:

• Python 3.6. Version 3.6.4 has been tested.
• PyTorch. Version 0.4.0 has been tested. Note that the GPU support is encouraged as it greatly boosts training efficiency.
• Other Python modules. Some other Python module dependencies are listed in requirements.txt, which can be easily installed with pip pip install -r requirements.txt

Testing the Project:

python impute.py -f data/data.npy -n data/network.pkl -o data/imputed_data.npy

Usage

Given an array of users' time series data and the social relationship among these users, this program can be used to replace the missing value in these time series data with reasonable values.

Input Format

The input files are expected to be two parts:

(1) data file: a numpy array (.npy) file which contains users' data shaped N * L * D, where N denotes the number of users, L denotes the sequence length and D denotes the number of channels at each time stamp. The missing data should be marked as -1 (or manually marked).

(2) social network file: a pickle file which contains the social network information formated as the adjacent list:

[[node0 's neighbors], [node1's neighbors],..., nodeN's neighbors]
e.g. [[1], [0,2,3], [1,3,4], [2]]

each node index is corresponding to the index of the row in the data array in (1).

See the sample data in the data directory.

Output Format

The program outputs to a file named imputed_data.npy which contains the data after imputation, i.e., the missing elements are replaced by reasonable values.

Main Script

The help of main script can be obtained by excuting command:

python impute.py -h
usage: impute.py [-h] [-f DATA_FILE] [-n SOCIAL_NETWORK] [-o OUTPUT_FILE]
                 [-m MISSING_MARKER] [-b BATCH_SIZE] [-e NUM_EPOCH]
                 [-s HIDDEN_SIZE] [-k DIM_MEMORY] [-l LEARNING_RATE]
                 [-d DROPOUT] [-r DECODER_LEARNING_RATIO] [-w WEIGHT_DECAY]
                 [--log]

optional arguments:
  -h, --help            show this help message and exit
  -f DATA_FILE, --data_file DATA_FILE
                        path of input file
  -n SOCIAL_NETWORK, --social_network SOCIAL_NETWORK
                        path of network file
  -o OUTPUT_FILE, --output_file OUTPUT_FILE
                        path of output file
  -m MISSING_MARKER, --missing_marker MISSING_MARKER
                        marker of missing elements, default value is -1
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        the number of samples in each batch, default value is
                        256
  -e NUM_EPOCH, --num_epoch NUM_EPOCH
                        number of epoch, default value is 200
  -s HIDDEN_SIZE, --hidden_size HIDDEN_SIZE
                        size of hidden feature in LSTM, default value is 32
  -k DIM_MEMORY, --dim_memory DIM_MEMORY
                        dimension of memory matrix, default value is 32
  -l LEARNING_RATE, --learning_rate LEARNING_RATE
  -d DROPOUT, --dropout DROPOUT
                        the dropout rate of output layers, default value is
                        0.8
  -r DECODER_LEARNING_RATIO, --decoder_learning_ratio DECODER_LEARNING_RATIO
                        ratio between the learning rate of decoder and
                        encoder, default value is 10
  -w WEIGHT_DECAY, --weight_decay WEIGHT_DECAY
  --log                 print log information, you can see the train loss in
                        each epoch

Reference

[1] Zongtao, L; Yang, Y; Wei, H; Zhongyi, T; Ning, L and Fei, W, 2019, How Do Your Neighbors Disclose Your Information: Social-Aware Time Series Imputation, In WWW, 2019

 @inproceedings{liu2019imputation, 
    title={How Do Your Neighbors Disclose Your Information: Social-Aware Time Series Imputation},
    author={Zongtao Liu and Yang Yang and Wei Huang and Zhongyi Tang and Ning Li and Fei Wu},
    booktitle={Proceedings of WWW},
    year={2019}
    }