Multi-directional Recurrent Neural Networks
- Datasets (GOOGLE.csv, GOOGLE_Missing.csv)
- These datasets are the example time-series datasets that can be used for testing M-RNN Architecture with Data_Loader.py
- Data_Loader.py
-
Using the Raw datasets, it extracts the features and time information.
-
It also divides training and testing sets for further experiments
-
It has two input parameters (1). train_rate: training / testing set ratio (2). missing_rate: the amount of introducing missingness
-
It has 4 outputs for each training and testing set (1). X: Original Feature (2). Z: Feature with Missing (3). M: Missing Matrix (4). T: Time Gap
- M_RNN.py
- Using the outputs of the Data_Loader.py, it imputes the missing features using M-RNN architecture
- It consists of Bi-directional GRU and MLP.
- The details of the M-RNN architecture can be found in the following link.
- https://arxiv.org/pdf/1711.08742.pdf
- M_RNN_Main.py
- Combine the above three components with MSE performance metric.