MTS2Graph introduces a new framework for interpreting multivariate time series data by extracting and clustering the input representative patterns that highly activate CNN neurons. This way, we identify each signal’s role and dependencies, considering all possible combinations of signals in the MTS input. Then, we construct a graph that captures the temporal relationship between the extracted patterns for each layer. An effective graph merging strategy finds the connection of each node to the previous layer’s nodes. Finally, a graph embedding algorithm generates new representations of the created interpretable time-series features. To evaluate the performance of our proposed framework, we run extensive experiments on eight datasets of Baydogan’s archive. The experiments indicate the benefit of our time-aware graph-based representation in MTS classification while enriching them with more interpretability.
Mustafa Baydogan's/The UEA multivariate time series classification archive MTS datasets can be obtained from: http://www.mustafabaydogan.com/files/viewdownload/20-data-sets/69-multivariate-time-series-classification-data-sets-in-matlab-format.html or contact Mustafa Baydogan
The code is divided as follows:
To run a model on a dataset with the default parameters, you should issue the following command:
To run the code using python Example_dataset.py -f <foldar_name_data> -d <dataset name>To set other hyperparameters
-c or --cluster_numbers: Sets the cluster numbers as a comma-separated list (e.g., 35,25,15).
-s or --segment_length: Sets the segment lengths (e.g.,10).
-a or --activation_threshold: Sets the activation threshold (e.g., 0.95).
-e or --embedding_size: Sets the embedding size (e.g., 100).
The FCN architecture used in MTS2Graph comprises three convolutional blocks with 32, 64, and 128 filters of varying lengths, ReLU activation functions, and batch normalization for regularization. For all datasets, we used 80% for training, 10% for validation, and 10% for test data. Each experiment is run for 500 epochs, with an embedding size of 100, a segment length of 10, an activation threshold of 0.95, and cluster numbers of 35,25,15.
The python packages needed are: