time-series-classification

Short and long time series classification via convolutional neural networks

In this project, we present a novel framework for time series classification, which is based on Gramian Angular Summation/Difference Fields and Markov Transition Fields (GAF-MTF), a recently published image feature extraction method. A convolutional neural network (CNN) was employed as the classifier. This framework enables the use of CNN to learn high-level features and classify time series. Its performance was evaluated on 16 standard datasets. Experiment results show that our framework outperforms or achieves the same level at least with the GAF-MTF+Tiled CNN framework on 14 of the 16 datasets. And it obtained competitive performance compared with other 8 representive approaches. Furthermore, we compared the performance of GAF-MTF feature with other 5 image features on a large-scale cough dataset. Results indicates that the GAF-MTF feature is not suitable for large-scale cough datasets while its competitive performance on the standard datasets.

Image features extraction

Short time series

Image features for short time series:

• GASF

Large-scale cough dataset

Image features for cough dataset:

• Comparision of the six image features:

  

CNN

• Framework for short time series classification:

Results

• short time series classification:

Appendix

Dataset information:

• Short: UCR Time Series Classification Archive
• Long: Cough dataset from Tongji Hospital

Software Links:

• Caffe: deep learning framework
• RASTAMAT

This project is partly motivated by @Zhiguang Wang, who is the author of "Imaging Time-Series to Improve Classification and Imputation". He provided me the source code to extract GASF-GADF-MTF features and pointed out that "The tiled CNN is not the best one and the TICA pre-training stage seems unnecessary". His advice helped us save a great deal of time. Thanks for his kindness and if you use this repository for GAF/MTF feature extraction, please cite the work in your publication:

@inproceedings{Wang:2015:ITI:2832747.2832798,
 author = {Wang, Zhiguang and Oates, Tim},
 title = {Imaging Time-series to Improve Classification and Imputation},
 booktitle = {Proceedings of the 24th International Conference on Artificial Intelligence},
 series = {IJCAI'15},
 year = {2015},
 isbn = {978-1-57735-738-4},
 location = {Buenos Aires, Argentina},
 pages = {3939--3945},
 numpages = {7},
 url = {http://dl.acm.org/citation.cfm?id=2832747.2832798},
 acmid = {2832798},
 publisher = {AAAI Press},
}

NOTE: The cough dataset used in this work can not be accessed now for some privacy issues!