String Kernels for Complex Time-Series: Counting Targets from Sensed Movement

Complex (imaginary) signals arise commonly in the field of communications in the form of time series in the complex space. In this work we propose a symbolic approach for such signals based on string kernels derived from a complex SAX representation and apply it to a challenging counting problem. Our approach, that we call cStrings, is within a Gaussian process regression framework and outperforms established Fourier transforms and complex kernels, achieving a correlation coefficient of 0.985 when predicting the number of targets sensed by a pulsed Doppler radar.