Abstract:

The classification of time series problems is widely applied across various fields. Existing classification metrics still face challenges when dealing with complex, noisy, and nonlinear time series. This paper introduces a method that combines complexity-invariant distance with complexity-invariant distance-based multidimensional scaling (CID-CIDMDS). This method demonstrates strong accuracy and robustness, and it does not depend on the length of the time series. The effectiveness of the method and the impact of sequence length are validated using time series generated from piecewise linear Lorenz maps, logistic maps, tent maps, and quadratic mapping models. The classification results are visually presented in a network format. The method's robustness is further confirmed through tests involving the addition of white noise, Gaussian noise, and impulse noise systems. Additionally, it is applied to the classification of stock market data.

Key words: time series, classification, CID, CIDMDS, network