Generalizing DTW to the multi-dimensional case requires an adaptive approach.

Area of Science:

• Data Mining
• Machine Learning
• Signal Processing

Background:

• Dynamic Time Warping (DTW) is widely adopted for time series data mining, especially in wearable device applications.
• Existing research primarily focuses on one-dimensional DTW, with multi-dimensional extensions often treated as equivalent or interchangeable.
• Practitioners commonly use dependent or independent warping for multi-dimensional DTW, assuming similar outcomes.

Purpose of the Study:

• To investigate the differences between dependent and independent warping methods in multi-dimensional Dynamic Time Warping.
• To demonstrate that these two common multi-dimensional DTW approaches can lead to distinct classification results.
• To introduce a principled rule for selecting the appropriate multi-dimensional DTW method for improved classification accuracy.

Main Methods:

• Comparative analysis of dependent and independent warping strategies for multi-dimensional DTW.
• Development of a predictive rule to guide the choice between the two methods.
• Extensive experimental evaluation on a large set of multi-dimensional time series classification tasks.

Main Results:

• Dependent and independent warping methods for multi-dimensional DTW are not equivalent and can produce different classifications.
• Neither dependent nor independent warping consistently outperforms the other across all datasets.
• The proposed rule effectively predicts which method to trust, ensuring accuracy comparable to or better than the best individual method.

Conclusions:

• The choice between dependent and independent warping in multi-dimensional DTW is critical and impacts classification outcomes.
• A simple, principled rule can reliably guide method selection, enhancing classification performance.
• This work provides a significant advancement in multi-dimensional time series classification by offering a robust and accurate approach.