A bag-of-features framework to classify time series.

Area of Science:

• Machine Learning
• Data Mining
• Pattern Recognition

Background:

• Time series classification is crucial for many applications.
• Nearest Neighbor (NN) with Dynamic Time Warping (DTW) is effective but has limitations.
• Existing feature-based methods struggle with local pattern translations and dilations.

Purpose of the Study:

• To present a novel framework for time series classification.
• To address the limitations of existing feature-based approaches.
• To improve classification accuracy by capturing local pattern variations.

Main Methods:

• Developed a Time Series Bag-of-Features (TSBF) framework.
• Utilized random subsequences of varying lengths and locations.
• Partitioned subsequences into intervals to capture local information.
• Employed a supervised learner integrating location information via class probability estimates.

Main Results:

• TSBF demonstrates effectiveness in handling local pattern warping, distinct from DTW.
• The framework successfully integrates location information into a compact codebook.
• Experimental results show TSBF outperforms NN classifiers and other alternatives on UCR benchmark datasets.

Conclusions:

• TSBF offers a robust feature-based approach for time series classification.
• The method effectively captures local series properties and handles warping.
• TSBF achieves superior performance compared to competitive methods on benchmark datasets.