A new time series classification approach based on recurrence quantification analysis and Gabor filter

This paper presents a new data stream preprocessing approach based on texture analysis from Recurrence Plot (RP). We adopt 12 real datasets related to sound recognition, signals processing, chemistry reactions and others. We consider data streams as discrete time series with a fixed number of observations. Classifier models were created using SVM and C5.0 algorithms with features extracted from RP using texture descriptors like to Gabor filter, GLCM and DCT-2D. Our analysis shows that the influence of stochastic-deterministic rate obtained by RQA on data stream reconstructed in phase space have motivated the preprocessing step to classification, outperforming another approach applied an audio stream that uses model's coefficients as features in around 16% of average accuracy and in around 67.66% the state-of-art algorithm that uses distance as measure.