Application of maximal information coefficient and affinity propagation to characterizing seismic time series associated with earthquakes

Abstract

Appropriate feature-based representations are significant for time series analysis and subsequent machine learning applications. A low-dimensional set of comprehensive features is instrumental to improving the efficiency and accuracy of classification. The main contribution of this work is to develop a new methodology to characterize seismic time series signals by extracting and selection statistical features from them. This methodology allows one to study earthquakes with much lower-dimensional, yet informative, datasets. In this work, a large number of unbiased features were generated from raw time series using the highly-comparative times series analysis (HCTSA) operation library. The similarity between each pair of features was represented by the measure of maximal information coefficient (MIC). MATLAB functions were implemented to compute the similarity matrix of the feature dataset generated by HCTSA. Affinity propagation (AP) was used for clustering similar features and selecting exemplary features from different clusters. These independent exemplary features were determined to characterize the original data. The process was applied to data from real earthquakes and a representation of reduced features was generated to characterize the original times series signals. Results showed that MIC reflected a reliable measure of general associations (similarities) between features, and features which were associated tended to be placed into the same cluster. The clustering results also showed that the average distance within a cluster was generally less than the distances between different clusters, which demonstrated that the selected exemplary features were relatively independent. This work opens a door to the use of feature-based datasets from seismic signals for model inversion and improved subsurface characterization.