DTW(dynamic time warping) is an early (1994, the paper is older than me) and classic algorithm in time series analysis. It actually borrows the idea of "dynamic programming" from the classical algorithm. Generally speaking, if we want to classify time series data, the experimental steps can be roughly divided into: data preprocessing (denoising or data enhancement), data characterization, and classifier selection (machine learning algorithm also needs to choose a suitable distance calculation method). Although the DTW algorithm also gives the path, I really can't figure out how to use the path, so I prefer to classify the DTW algorithm as a distance calculation method.

The first part of the introduction is not introduced. Directly introduce the second part: dynamic time warping.

The author first mentioned the successful application of dtw algorithm in the field of speech recognition-researchers matched the pronunciation of a word in reality (actually a time series) with the pronunciation of words in the template library one by one. How to measure the matching degree?