Efficient Kernel-Based Subsequence Search for Enabling Health Monitoring Services in IoT-Based Home Setting.

Area of Science:

• Data Science
• Machine Learning
• Signal Processing

Background:

• Subsequence search in data streams is crucial for pattern recognition.
• Dynamic Time Warping (DTW) is effective but computationally expensive.
• Existing methods struggle with the computational burden of DTW in streaming data.

Purpose of the Study:

• To develop an efficient kernel-based approach to approximate DTW for subsequence search.
• To enable the analysis of streaming data from wearable sensors with reduced computational load.
• To address the challenge of comparing time-series of different lengths using kernel methods.

Main Methods:

• Learning a kernel function that approximates DTW.
• Employing a feature embedding technique to represent time-series as fixed-length vectors.
• Utilizing DTW between time-series and randomly chosen basis series for vector components.

Main Results:

• The proposed kernel approach significantly reduces computational costs compared to traditional DTW.
• A slight decrease in accuracy was observed, which is offset by computational gains.
• Validation on benchmark datasets and a real-life elderly self-rehabilitation application.

Conclusions:

• The kernel approximation of DTW offers an efficient alternative for subsequence search in data streams.
• This method is particularly beneficial for real-time analysis of sensor data.
• The approach balances accuracy with substantial computational efficiency.