Incremental and Adaptive Feature Exploration over Time Series Stream

DAVID Lab
Unversity of Versaille Saint-Quentin (UVSQ)
Université Paris-Saclay

A demonstration video can be found here -> Demo Video [1]

You can find the paper here -> IEEE Big Data 2019

By adopting the concepts Shapelet [2] and Matrix Profile [3], we conduct the first attempt to extract the incremental and adaptive features from Time Series Stream:

• For data source with stable concept, learning model will be updated incrementally;
• For data source with Concept Drift, we extract the adaptive Shapelets under the most recent concept;
• An elastic caching mechanism allows to tackle the infinite TS instances in streaming context.

Configurations (demo)

• Input File: the name should be end with "Train.csv"
• dataset_folder: in each file, change the location of the datasets in the background. The selected input file will be saved/uploaded into this folder.
• Data Augmentation: refer to preprocessing/TS_stream_preprocess.py. As Shapelet-based methods (e.g., SMAP) are noise resistant, we put randomly the noise of random durations into the original TS data to augment the data volume.

Web application (demo)

• ISETS_webapp.py: main program, a web application based on Flask and Bokeh
• ISETS_webbackend.py: the program for adaptive shapelet extraction and Concept Drift detection
• draw_adaptive_shapelets.py: show the adaptive shapelets in the web interface
• draw_TS_Stream.py: show in real time the input TS instances in the stream

Core algorithms

• utils/: the repository which contains the basic file operations and similairty measure functions
• memory_block.py: the caching mechanism including the computation of Matrix Profile for cached instance
• SMAP_block.py: Shapelet extraction on MAtrix Profile
• evaluation_block.py: the loss computation and the Concept Drift detection on TS Stream
• adaptive_features/adaptive_features.py: Concept Drift detection and adaptive feature extraction
• ISMAP/ISAMP.py: incremental Shapelet extraction on MAtrix Profile

Main contributions:

1. A novel strategy to evaluate Shapelet, which shows the first attempt of transferring the techniques in Time Series community to Data Stream community

   ​

   ​ Figure 1. Shapelet Evaluation by a loss-smoothed approach.

2. Test-then-Train Stategy: The novel strategy, not only accelerates the incremental Shapelet extraction in stable-concept context, but also helps with detecting Concept Drift in streaming context.

   

3. Elastic Caching Mechanism in Streaming context

   

4. Scalability & Explainability

   Based on SMAP (Shapelet extraction on MAtrix Profile) [4], our system is capable of being distributed in Spark environment. In addition, apart from the interpretability provided by Shapelet features, the system shows a strong explainability for Shapelet Extraction and Shapelet updating process.

5. Traceability of extracted features
   
   
   
   

Citation

If you find this repository useful in your research, please consider citing the following paper:

@inproceedings{zuo2019incremental,
  title={Incremental and Adaptive Feature Exploration over Time Series Stream},
  author={Zuo, Jingwei and Zeitouni, Karine and Taher, Yehia},
  booktitle={2019 IEEE International Conference on Big Data (Big Data)},
  pages={593--602},
  year={2019},
  organization={IEEE}
}

[1] J. Zuo, K. Zeitouni, and Y. Taher, “ISETS: Incremental Shapelet Extraction from Time Series Stream“，(demo paper) ECML-PKDD’19

[2] L. Ye and E. Keogh, “Time series shapelets: A New Primitive for Data Mining,” in Proc. KDD 2009

[3] C. Yeh, Y. Zhu, L. Ulanova, and N. Begum, “Matrix Profile I: All Pairs Similarity Joins for Time Series: A Unifying View that Includes Motifs, Discords and Shapelets,” IEEE ICDM 2016

[4] J. Zuo, K. Zeitouni, and Y. Taher, “Exploring interpretable features for large time series with se4tec,” in Proc. EDBT 2019